Add experimental timeline video reference modes

This commit is contained in:
ukr8b3g-cmyk
2026-09-22 03:02:57 +09:00
parent 085943ad9c
commit 178efb6b21
26 changed files with 5607 additions and 81 deletions
+4
View File
@@ -44,3 +44,7 @@ examples/workflows/*
# Current loader migration instructions. # Current loader migration instructions.
!docs/LOADER_PERSISTENCE_REPAIR.md !docs/LOADER_PERSISTENCE_REPAIR.md
!docs/TIMELINE_VIDEO_EXPERIMENTAL.md
# Local updater receipt; never distribute it as package content.
.h3-timeline-experimental.json
+7 -8
View File
@@ -57,16 +57,9 @@ Current release-hygiene baseline: O1 and AUDIO-R1 are PASS; the public surface i
- Hi-Res Fix is not part of the new standard workflow. Preserve its historical implementation until cleanup is approved; use Second Pass as the external latent-processor bridge. - Hi-Res Fix is not part of the new standard workflow. Preserve its historical implementation until cleanup is approved; use Second Pass as the external latent-processor bridge.
- Keep P1a First Block Cache and P2 boundary-aware Sol-Attn `Deferred`. Keep A5 and A8b `HOLD` under their recorded release contracts. Issue #13 reporter-exact R3 remains the highest Continuation Quality investigation when its required workflow is available. - Keep P1a First Block Cache and P2 boundary-aware Sol-Attn `Deferred`. Keep A5 and A8b `HOLD` under their recorded release contracts. Issue #13 reporter-exact R3 remains the highest Continuation Quality investigation when its required workflow is available.
## H3 completion review handoff
- After every implementation or Gate is complete and its local validation report is ready, send that self-contained completion report to the pinned ChatGPT task titled `H3情報チェック` before starting the next phase.
- The fixed current destination is task ID `6a94b91e-af7c-83e8-be18-ddc504f690c4` (pinned index 3). A second older task has the same title; do not send to the unpinned duplicate.
- Wait for and read the review reply, then bring its conclusions and requested corrections back into the active Continuum task as feedback. Treat the reply as review input under the user's authority; preserve the repository rules and request approval where a proposed change expands scope.
- Do not duplicate a handoff that the user explicitly says they already completed manually. If the destination is unavailable or the reply requires user judgment, stop and report that condition instead of silently skipping the review.
## A8b and A5 HOLD release contracts ## A8b and A5 HOLD release contracts
- 2026-09-05 explicit user exception ("許可、許可。"): Issue13 R3B may run one isolated Video-only weak tapered context-noise A/B experiment before reporter-exact R3. This waives only the reporter-exact/independence prerequisite for that diagnostic. Keep Production/source runtime unchanged, Audio/masks/depth/Seed/SIGMAS/grouping unchanged, default OFF bit-exact, and pass CPU/Shadow/Replay before GPU. Restore clean exported prefix after the diagnostic perturbation. No combined interventions, A5 work, public controls, or Production promotion are authorized. General A8b and Release HOLD remain; GPU-audit sharing is limited to the fixed H3情報チェック task. - 2026-09-05 explicit user exception ("許可、許可。"): Issue13 R3B may run one isolated Video-only weak tapered context-noise A/B experiment before reporter-exact R3. This waives only the reporter-exact/independence prerequisite for that diagnostic. Keep Production/source runtime unchanged, Audio/masks/depth/Seed/SIGMAS/grouping unchanged, default OFF bit-exact, and pass CPU/Shadow/Replay before GPU. Restore clean exported prefix after the diagnostic perturbation. No combined interventions, A5 work, public controls, or Production promotion are authorized. General A8b and Release HOLD remain. Do not automatically send cross-task audit reports; sending requires a fresh explicit user request.
- A8a is the accepted read-only baseline: full pytest `1041`, Manifest `242/242`, every observer record has `execution_applied=false`, and Production/Sampling/Run Storage/State/Session/Assembly parity is preserved. - A8a is the accepted read-only baseline: full pytest `1041`, Manifest `242/242`, every observer record has `execution_applied=false`, and Production/Sampling/Run Storage/State/Session/Assembly parity is preserved.
- Keep A8b execution policy on HOLD until one intervention type is specified and approved. Do not change context, transport, mask, latent, or conditioning in combination. Seed, SIGMAS, and physical grouping remain unchanged unless separately approved. - Keep A8b execution policy on HOLD until one intervention type is specified and approved. Do not change context, transport, mask, latent, or conditioning in combination. Seed, SIGMAS, and physical grouping remain unchanged unless separately approved.
@@ -107,3 +100,9 @@ GPU checks must cover Sage only, Sage+Sol, Sage+Spectrum, and Sage+Sol+Spectrum
- Keep H3EasyLoadImage and Enable Image; delegate inherited mode accessors and leave Core widget collections, drawing and serialization intact. - Keep H3EasyLoadImage and Enable Image; delegate inherited mode accessors and leave Core widget collections, drawing and serialization intact.
- New official workflows use Core LoadAudio and Core LoadVideo -> H3ContinuumVideoAdapter. The adapter has no Enable/file/upload UI and reuses the existing frame-rate conversion. Preserve the original Audio/Video IDs as deprecated compatibility nodes; never infer lost filenames or saved OFF states. - New official workflows use Core LoadAudio and Core LoadVideo -> H3ContinuumVideoAdapter. The adapter has no Enable/file/upload UI and reuses the existing frame-rate conversion. Preserve the original Audio/Video IDs as deprecated compatibility nodes; never infer lost filenames or saved OFF states.
- The approved export exception adds H3ContinuumVideoAdapter to the existing nine-node mapping. Do not change Main Sampler, Sampling, external wrappers, Decode Cache or Run Storage semantics for this repair. - The approved export exception adds H3ContinuumVideoAdapter to the existing nine-node mapping. Do not change Main Sampler, Sampling, external wrappers, Decode Cache or Run Storage semantics for this repair.
## User-approved Timeline Video experiment (2026-09-22)
- Append one optional `Video Reference Mode` to V38. Missing mode remains `Repeat Reference`; `Follow Timeline` is opt-in and uses each physical group's visible output range without consuming continuation context twice.
- Preserve grouping, Seed, SIGMAS, Audio, masks, Run Storage schema, Terminal Merge ownership, and the two official V3.8X2 workflow files. Experimental Follow/Repeat templates remain separate comparison assets.
- Local acceptance passed: full CPU `1424 passed / 1 skipped`, browser save/reload, Follow/Repeat GPU functional runs, A/B mode separation, and Manifest integrity. Keep the feature Experimental; do not describe it as exact motion copying or a long-input RAM optimization.
+10
View File
@@ -1,5 +1,15 @@
# Changelog # Changelog
## Experimental on main — Unified Timeline Video (2026-09-22)
- Renamed the existing V3.8 video IMAGE input to `Timeline Video Frames` while preserving its internal key and saved links.
- Added one appended optional `Video Reference Mode`: `Repeat Reference` remains the compatibility default; `Follow Timeline` selects the source interval belonging to each physical output group.
- Follow mode excludes continuation-prefix frames from source consumption, records each physical slice in the plan, uses slice-specific Video VAE/cache identity, and continues without video reference after source exhaustion.
- Kept Sampling, Seed, SIGMAS, Audio, masks, continuation grouping, Terminal Merge ownership, Run Storage schema, and the two official V3.8X2 workflows unchanged. Added separate Follow/Repeat comparison workflows.
- Synchronized the lightweight runtime verifier and README public-surface wording with the already-exported ten-node V3.8X2 surface, including `H3ContinuumVideoAdapter`.
- Acceptance: full CPU `1424 passed / 1 skipped / 0 failed`; browser save/reload PASS; Follow and Repeat `2 x 5 s` GPU runs PASS at 704x416/24fps with synchronized 10-second audio/video and no OOM, NaN, allocation failure, or crash.
- This remains Experimental: it is soft H3 conditioning, not exact motion copying, and it does not reduce the RAM needed to hold a long decoded IMAGE batch.
## Unreleased — Loader persistence repair (2026-09-21) ## Unreleased — Loader persistence repair (2026-09-21)
- Repair Issue #23 by delegating inherited native mode accessors; preserve Enable Image and remove draw/save-time widget collection replacement. - Repair Issue #23 by delegating inherited native mode accessors; preserve Enable Image and remove draw/save-time widget collection replacement.
+22 -14
View File
@@ -1,10 +1,10 @@
dca65fb1393e67f59c6f81f9411e3e745eb8f1097e9a2adf7ffc6f4e3901ca9f __init__.py dca65fb1393e67f59c6f81f9411e3e745eb8f1097e9a2adf7ffc6f4e3901ca9f __init__.py
a9fd5a4c5624e17148656abed0ba41a72c72cac73e7f6f36d2816ad92962f867 .comfyignore bb6f7d1227f7c0cc4e134ccc0e037c8c4fe4af63d8c990ac00b7af2279695871 .comfyignore
3d681ded5d4c50cc20c2382e29fdc98d8a48e76d552c41ad5a195d941e66e612 .gitattributes 3d681ded5d4c50cc20c2382e29fdc98d8a48e76d552c41ad5a195d941e66e612 .gitattributes
8296f0afc7be47171e439c02fb2988cfeb0df904acbe3fc4f507f426d709157e .github/workflows/tests.yml 8296f0afc7be47171e439c02fb2988cfeb0df904acbe3fc4f507f426d709157e .github/workflows/tests.yml
0a2f69d4cec89d440f5d5aa1ec68a7c9a741382d313f80dcddf96d13f70b3e68 .gitignore 0a2f69d4cec89d440f5d5aa1ec68a7c9a741382d313f80dcddf96d13f70b3e68 .gitignore
3e7a1ec3380be2a222c8a3e9b8a6bb184d3a760478957f48186438f15235bdba branch_provenance.py 3e7a1ec3380be2a222c8a3e9b8a6bb184d3a760478957f48186438f15235bdba branch_provenance.py
d70402478f19f6487eed08957a048a8c248a9291be8e62c0e7ade69da82fd393 CHANGELOG.md ad340adc9506ab069a83f586d954fbd19f44df4e18a5346d5a8923a489d3cb85 CHANGELOG.md
c1ed3a0342730450958fb62fd21854a876991f3a4aefb24e5383b0c73cd2312d compatibility.py c1ed3a0342730450958fb62fd21854a876991f3a4aefb24e5383b0c73cd2312d compatibility.py
4d62508f900af9f93ec320afb900c27cc9a300d1e0c0a94714258498c5ede5fc conditioning.py 4d62508f900af9f93ec320afb900c27cc9a300d1e0c0a94714258498c5ede5fc conditioning.py
294251c40b1be2e296c5e5f627b68091d087bdb39179e3a0c3cebd2ee8a344f5 constants.py 294251c40b1be2e296c5e5f627b68091d087bdb39179e3a0c3cebd2ee8a344f5 constants.py
@@ -118,12 +118,12 @@ e2b72149b0433e95c13659cc9e742fe38523e67d4f4a2fdb08aed1943fecf653 nodes.py
b470819f16fc179654dc1600721a63f0a5a14e777f4c004d574b499da6886ba5 public_api.py b470819f16fc179654dc1600721a63f0a5a14e777f4c004d574b499da6886ba5 public_api.py
15a3925218e0c5f968795d7d11e4783c3c6868917447274bc700328851b3fbfb pyproject.toml 15a3925218e0c5f968795d7d11e4783c3c6868917447274bc700328851b3fbfb pyproject.toml
94fdc2d545b2d4077a4d70229fd22ba3ccac21d49d1097d2873f76d04e9a3c7a pytest.ini 94fdc2d545b2d4077a4d70229fd22ba3ccac21d49d1097d2873f76d04e9a3c7a pytest.ini
4a0507ce6183e099265d5a34a38e43da89515af8eb680f5a48abc8aecc724d63 README_JA.md 6f108a67967dd686011bf031b8176373f6ab80635f7e01e3e935e8585b804fc9 README_JA.md
9d4eefcc97e74d4c1c8ab1be612b50eb99a8a4f8db7c2134ca85a9e67bf2568a README.md 77cdaadd32f9d722f6b05abc196cbdb9dded32387493664620a30fa405429743 README.md
44127adcfa30ee84bff9cd4b2d0c695355209550e76982d374122008728aa5bc reference_audio.py 44127adcfa30ee84bff9cd4b2d0c695355209550e76982d374122008728aa5bc reference_audio.py
5cb9bfd6810d3a93a637451980a4cc5c1436e5b84cfd1c7a111cd192e876351b reference_video.py 5cb9bfd6810d3a93a637451980a4cc5c1436e5b84cfd1c7a111cd192e876351b reference_video.py
c7b14ed6f8cfc2a1f0a21db4fe56fba75a93ffcb0be0c22bbd2e691c45cd0e2d reference.py c7b14ed6f8cfc2a1f0a21db4fe56fba75a93ffcb0be0c22bbd2e691c45cd0e2d reference.py
3157b7b0d307d05d109b940021b0ca09a2ab2680c4c362584399837e9434d728 REGISTRY_MANIFEST.sha256 ed892d90b644050bdf68f860c3b5bcdef6c59b565257275bbe7b07238ab1f630 REGISTRY_MANIFEST.sha256
467edb2c188d024c2392aafe7423b3f835e16b10f570da76e2f7dda8a962c1c5 requirements.txt 467edb2c188d024c2392aafe7423b3f835e16b10f570da76e2f7dda8a962c1c5 requirements.txt
8bfa113a2ec11325619be300b4d2b53a157b1d2c1ef7855057a4b3bbd05aeef3 run_storage.py 8bfa113a2ec11325619be300b4d2b53a157b1d2c1ef7855057a4b3bbd05aeef3 run_storage.py
6056de6c0a5322a8a51ba5a114538dca2ef67788cd1afba10fedd483b95f6afa state_io.py 6056de6c0a5322a8a51ba5a114538dca2ef67788cd1afba10fedd483b95f6afa state_io.py
@@ -228,15 +228,15 @@ e813b5d21af2963a325862acd344b33caa124d0cd3940a940c58118cf27742e0 tests/test_v38
c9577e9b2d430c9435a287bf679a920020192d51f780e02b47f286e2e4b6a615 tests/test_v38_easy_mode.py c9577e9b2d430c9435a287bf679a920020192d51f780e02b47f286e2e4b6a615 tests/test_v38_easy_mode.py
bb5a1530c890d357dc48ea74039340d746e876e8f85e06d4d9a19369f8ff5961 tests/test_v38_hover_help.py bb5a1530c890d357dc48ea74039340d746e876e8f85e06d4d9a19369f8ff5961 tests/test_v38_hover_help.py
98eb59407741dd187dc22d94b4f433f341294d05cc38106fde15b580b1465001 tests/test_v38_memory_action_policy.py 98eb59407741dd187dc22d94b4f433f341294d05cc38106fde15b580b1465001 tests/test_v38_memory_action_policy.py
4338c932fc8a00328c3145b276e0f58a783c721a683b43aa6c24337b05a878ab tests/test_v38_multi_reference_audio.py 8edaae5f1ca93925e9269c9a4447c97edf5958395f08659e61a09e8d7efa088a tests/test_v38_multi_reference_audio.py
623d528a02b6712cba28c374ba3ad979bc84e2ac5318640692589a04ba6ab15f tests/test_v38_open_integration_contract.py 623d528a02b6712cba28c374ba3ad979bc84e2ac5318640692589a04ba6ab15f tests/test_v38_open_integration_contract.py
12337a491fa67b38b32aaeb16e0b209393960dbc46db2a8bf968b8c3902855ef tests/test_v38_packed_row_planner.py 12337a491fa67b38b32aaeb16e0b209393960dbc46db2a8bf968b8c3902855ef tests/test_v38_packed_row_planner.py
abece11f7efa1f09a4ba4ab729708f460277128ac590c1e16485dd601b1c499c tests/test_v38_public_surface_cleanup.py 71a042353883ec6c86bbab2027c89087d479a419e9d4a1801d6852904c980e04 tests/test_v38_public_surface_cleanup.py
927348e89480b2f3b2fc2dcffa1398931d20e8812d94374ad05b374b453b7814 tests/test_v38_public_workflow.py 927348e89480b2f3b2fc2dcffa1398931d20e8812d94374ad05b374b453b7814 tests/test_v38_public_workflow.py
14ff6561a127287d0a20887790a978612ef9d291f1d3be7d5e85b37b2c74760e tests/test_v38_ref_encode_cache.py 14ff6561a127287d0a20887790a978612ef9d291f1d3be7d5e85b37b2c74760e tests/test_v38_ref_encode_cache.py
8b376f45261d045b584f02b4d14cd70c2d6269b9c814d4e62b5c612f6e30e36f tests/test_v38_refine_target.py 8b376f45261d045b584f02b4d14cd70c2d6269b9c814d4e62b5c612f6e30e36f tests/test_v38_refine_target.py
a67a52da8828bdff5d349c40d48c102271d7c9e388120565f50e76941b01f91c tests/test_v38_reliability.py 03eb5bf6d845b22334fcc800d81b419f1c8d8593fa65a1e0460144456ff2c8ba tests/test_v38_reliability.py
d512b0b277c7f1e53c3512b00055e237468e2f0f904639dfa4307ef28d39665a tests/test_v38_resolution_preset.py c7713198c14316d62425f15b4273d5998cc284d32da3babd6b296a496040721b tests/test_v38_resolution_preset.py
2ac395009e49f9554af8d1a251d93fc5d5e7070bf80f5a0b7b3f7e00cde15077 tests/test_v38_review_cache.py 2ac395009e49f9554af8d1a251d93fc5d5e7070bf80f5a0b7b3f7e00cde15077 tests/test_v38_review_cache.py
72a0e3cb334049b66dd36f52c826931067c242fa63557933fd667138b38d6e8d tests/test_v38_review_control.py 72a0e3cb334049b66dd36f52c826931067c242fa63557933fd667138b38d6e8d tests/test_v38_review_control.py
f62daa235117a3041c66ebb17adedb13c0452c3a2fc0acdd51006d7aa0078a9b tests/test_v38_review_execution_cap.py f62daa235117a3041c66ebb17adedb13c0452c3a2fc0acdd51006d7aa0078a9b tests/test_v38_review_execution_cap.py
@@ -300,7 +300,7 @@ cc247c01a01efaf594c24c7ce36ca3357c092fc28dd742c7577093ee9d5efdef tools/v38_reso
ad8609a45ba818fb74245b30cb0a3a221d6b345c20671092cafd714d8b07aaea tools/v38_review_phase_f_gpu_gate_runner.py ad8609a45ba818fb74245b30cb0a3a221d6b345c20671092cafd714d8b07aaea tools/v38_review_phase_f_gpu_gate_runner.py
60e9bf2c7c40ef9dd71f08ec2ec8c25f7ae9b08e74c3b7be5964e1138848ae82 tools/v38_review_phase_f_recovery_runner.py 60e9bf2c7c40ef9dd71f08ec2ec8c25f7ae9b08e74c3b7be5964e1138848ae82 tools/v38_review_phase_f_recovery_runner.py
a18136a4decfeef344635bbe9bb18502231c2503b034ce1d21cc2e5e77cc5e2f tools/validate.ps1 a18136a4decfeef344635bbe9bb18502231c2503b034ce1d21cc2e5e77cc5e2f tools/validate.ps1
087f4f727e9f76bb2b0061e26e0b49065370aa352a1ebebc7f253774b3be22ed tools/verify_runtime.py deeae570772c1a32dbb356aa30c8ef473b9334a0a36b71c0cccba12203f76415 tools/verify_runtime.py
ffd61f3f31c9b6430e1169007ee906c8f5000eba63ee44e2976ad4f5f1112671 v2/__init__.py ffd61f3f31c9b6430e1169007ee906c8f5000eba63ee44e2976ad4f5f1112671 v2/__init__.py
2003902c87bba09cfa143f5a9e4248f3a78c0fb544449a75fd40ba2cacb35655 v2/context_diagnostics.py 2003902c87bba09cfa143f5a9e4248f3a78c0fb544449a75fd40ba2cacb35655 v2/context_diagnostics.py
cde27db19b677e318cc4dfb4235be7a11429c19945f4083052a91fe117e74c82 v2/decoder.py cde27db19b677e318cc4dfb4235be7a11429c19945f4083052a91fe117e74c82 v2/decoder.py
@@ -313,7 +313,7 @@ dc2b901789e4b024dbcbc4024c63eb8cbc564d77003ce8c35c0584851ba6dfbd v2/prompts.py
2fd811794c4cc0f2ce050250b80484876454926b9f20fce6af3a97d728e5acdf v2/seam_guard.py 2fd811794c4cc0f2ce050250b80484876454926b9f20fce6af3a97d728e5acdf v2/seam_guard.py
be0d8585fa57e0ccd6d02577bbc7299629e5bfe20e186747f571fccc98b67d7e v2/seam_types.py be0d8585fa57e0ccd6d02577bbc7299629e5bfe20e186747f571fccc98b67d7e v2/seam_types.py
12cfcc0c71ca087fe39a5ce2922959a62b2bb311bfe664f05c4480affce03708 v2/seeds.py 12cfcc0c71ca087fe39a5ce2922959a62b2bb311bfe664f05c4480affce03708 v2/seeds.py
ff23b9f49b05b1e000110ab4eb211942f0aa0d376f35694e80c5df25047a77c7 v2/sequence.py fe7df9efea15b157c158046f64328b3dce3cfbd35cf180573ce4ec771ddcea8c v2/sequence.py
13213e7709fd02bb810af3c18e4cd7d010f8cdfb121a4b99866a9f2333331089 v2/session_io.py 13213e7709fd02bb810af3c18e4cd7d010f8cdfb121a4b99866a9f2333331089 v2/session_io.py
3ea23dbce6fcd217b6e81235855479b9b16fed7a21693cf156f9014555c4fe13 v2/session.py 3ea23dbce6fcd217b6e81235855479b9b16fed7a21693cf156f9014555c4fe13 v2/session.py
78c830cdf49034c6e3b2cd3da5914d5cfc5638eb63ec6dbd1bab13c50d6d4caf v3/__init__.py 78c830cdf49034c6e3b2cd3da5914d5cfc5638eb63ec6dbd1bab13c50d6d4caf v3/__init__.py
@@ -321,7 +321,7 @@ ea2922567df90436dd6fa21696833a49dd0cbfeee142a5fedc3a7c303824d473 v3/adaptive_co
b3dc61503d35d9ef0a0dc7fed1c9b2fdc00fcb2072d4e39f64ae0826f9ec7dca v3/assembly_v35.py b3dc61503d35d9ef0a0dc7fed1c9b2fdc00fcb2072d4e39f64ae0826f9ec7dca v3/assembly_v35.py
7815230e0a9540bc2edeaf818e940b7a1fdd817fd217ba8a4f9cb21aaf6a308e v3/assembly.py 7815230e0a9540bc2edeaf818e940b7a1fdd817fd217ba8a4f9cb21aaf6a308e v3/assembly.py
3f62fdb35e804f42d7cbf0828b4e87e69a2b99246c3f7467fe4fa698f1d803c9 v3/conditioning_bridge_nodes.py 3f62fdb35e804f42d7cbf0828b4e87e69a2b99246c3f7467fe4fa698f1d803c9 v3/conditioning_bridge_nodes.py
aa044669259df440a384b5e839f75fa9665790da0c210a427375fdd5c76af779 v3/driving_nodes.py 82238eff890574fcf977dc3d2bc092287f63710a47bfc0f6692c272814dbd295 v3/driving_nodes.py
147a1d3fc5e09b323a4c3d13d30e432410389bae52b58330462375b9f307b791 v3/easy_audio_nodes.py 147a1d3fc5e09b323a4c3d13d30e432410389bae52b58330462375b9f307b791 v3/easy_audio_nodes.py
64b3e54b42e5156642f9957ccc374fd190d9fc7145637984e702534fb1bf80b4 v3/easy_image_nodes.py 64b3e54b42e5156642f9957ccc374fd190d9fc7145637984e702534fb1bf80b4 v3/easy_image_nodes.py
ac859cb1a6e23b8ceae300d818c24fe1346b649a0d663017a9b3ac64aa96fca3 v3/easy_nodes.py ac859cb1a6e23b8ceae300d818c24fe1346b649a0d663017a9b3ac64aa96fca3 v3/easy_nodes.py
@@ -365,10 +365,18 @@ fefafd4550511d80bd85ade46a0476f0cf3de13ddb94a908dff6b175f2a0e199 web/easy_bypas
9db0aa94f634578a962d5db49e5c70356040a59df968e3a2e531b327a623081f web/easy_load_image.js 9db0aa94f634578a962d5db49e5c70356040a59df968e3a2e531b327a623081f web/easy_load_image.js
35090816ba7470a7255eb85d003b7e79f8a0065982dacf98a18f41c8ebff655d web/easy_load_video.js 35090816ba7470a7255eb85d003b7e79f8a0065982dacf98a18f41c8ebff655d web/easy_load_video.js
0c68d13a1f8beb7431a32b471f8b0612c42fcc4bf72dd97611a025d158008662 web/easy_v38.js 0c68d13a1f8beb7431a32b471f8b0612c42fcc4bf72dd97611a025d158008662 web/easy_v38.js
ec591d1b8f7587594397b995b9584809435f10fa3de6bf8f8df204c34809aa43 web/project_id.js bb7e29359e1fe26fd70d22f6bc70eb8c8858f68ab954ba39ad3670f2cdeeae7d web/project_id.js
ecdcd27528f6f1e80f758c608cbc42f747163ea4c39abd77576b91c627423d12 web/reference_audio_ui.js ae59273f163a116ece418957c557c4fd8c3285e3c60814679b9361d2e6e6fa02 web/reference_audio_ui.js
46c9f75c9462bf26300332448f4431cca801f38964ca0a49bc2c50a88a7cc40f web/reference_image_ui.js 46c9f75c9462bf26300332448f4431cca801f38964ca0a49bc2c50a88a7cc40f web/reference_image_ui.js
26b13614ba0a5fd8b192a106c9503b2f4b6fe7f9268a0fcb0f544979a49cab34 v3/video_adapter_nodes.py 26b13614ba0a5fd8b192a106c9503b2f4b6fe7f9268a0fcb0f544979a49cab34 v3/video_adapter_nodes.py
a9da2d25e9fc9b87b249274d7bf2ab061d14f5d1e9c482b69d795ba511cdace1 tests/test_loader_store_roundtrip.py a9da2d25e9fc9b87b249274d7bf2ab061d14f5d1e9c482b69d795ba511cdace1 tests/test_loader_store_roundtrip.py
0e02f7db41458fab91e6c3395bdb864d1c033923a3874ff4082d5f85b8230d98 tests/test_video_adapter.py 0e02f7db41458fab91e6c3395bdb864d1c033923a3874ff4082d5f85b8230d98 tests/test_video_adapter.py
77c225d1e49e06d00b04acf811177c9a4317e792d4bfe92df539de99420c3563 docs/LOADER_PERSISTENCE_REPAIR.md 77c225d1e49e06d00b04acf811177c9a4317e792d4bfe92df539de99420c3563 docs/LOADER_PERSISTENCE_REPAIR.md
b935ad3bb729948bb3890ce967b0ecbef737a5e51ddfc47a0060fc20aa78dcda video_reference_modes.py
bf41bf08dd0504eb753fa5781e6c2b8c1285758ad4bb7309975bb9231d5f8e41 web/video_reference_mode_ui.js
4c33d223a77c78e5c74946aa43f537031fd5aa415388d46fed975f9b07d53c3e docs/TIMELINE_VIDEO_EXPERIMENTAL.md
b07b941082a947d9e938ff399ec8b77fcd29c1a222e0e6b0dedf5d8e7482ce84 tests/test_video_reference_modes_experimental.py
171f0b5596724926d1d89fd12a46434f6010f12e651d19865b9b6160a1e9fff6 examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.json
e61e45f218609a7ec845d2e8762dfd1efed6c51a398061066d1e51fc93556663 examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.zip
7d6754cfc75599e07046b1949c1723a56f57871ecd5abc9543fcf8b34bca8c04 examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.json
9b60c8495e3d0e32c2bcd609080342ae19d0fd4c7c0662d5fe32e7fad4a9b41c examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.zip
+7
View File
@@ -1,5 +1,12 @@
# Project State # Project State
## Timeline Video Experimental — local acceptance PASS, main publication candidate (2026-09-22)
- Based on main `085943ad9cb8023a2042d4d731c83e638361c06e`. The existing `reference_video_1` IMAGE socket is displayed as `Timeline Video Frames`; one optional `video_reference_mode` widget is appended. Missing values default to `Repeat Reference`; `Follow Timeline` is opt-in.
- Follow selects each physical group's visible output interval, excludes continuation-prefix frames, keeps Terminal Merge ownership, and continues without video reference after source exhaustion. Per-slice identity prevents conditioning/VAE cache reuse across different intervals. Sampling, Seed, SIGMAS, Audio, masks, Run Storage schema, and official V3.8X2 workflow bytes remain unchanged.
- Local acceptance: full CPU `1424 passed / 1 skipped / 0 failed`; browser save/reload PASS; Follow and Repeat `2 x 5 s` GPU functional runs PASS at 704x416/24fps with synchronized 10-second AV; A/B diverges in Chunk 2; no OOM, NaN, allocation failure, or crash. Manifest/Registry Manifest matched before publication preparation.
- The feature remains Experimental. It is not exact motion copying, a variable-duration chunk system, or a long-input RAM optimization. Separate Follow/Repeat comparison workflows are source assets; Release/tag and Registry publication remain separate operations.
## Loader repair full-suite follow-up (2026-09-21) ## Loader repair full-suite follow-up (2026-09-21)
- Main repair `f8eb40d8b44e54386bcf3c5197c48f0f71daf6df` passed the 46 focused tests. Its first full CI returned 1376 passed, 3 skipped, and one stale pre-migration workflow fingerprint failure in `test_v38x2_decode_cache_integration.py`. - Main repair `f8eb40d8b44e54386bcf3c5197c48f0f71daf6df` passed the 46 focused tests. Its first full CI returned 1376 passed, 3 skipped, and one stale pre-migration workflow fingerprint failure in `test_v38x2_decode_cache_integration.py`.
+48 -20
View File
@@ -2,7 +2,7 @@
## Loader persistence repair on main (2026-09-21, after 3.8.3) ## Loader persistence repair on main (2026-09-21, after 3.8.3)
The current paired V3.8X2 workflows now use **Core Load Audio** and **Core Load Video -> H3 Continuum Video Adapter**. The adapter has Force Rate (default 24) and IMAGE/AUDIO outputs, but **no Enable switch or file selector**. Core owns video selection/upload/save; the established duration-preserving frame selection is reused and audio is passed unchanged. To disable Video Guide, bypass the adapter or the whole video-input group; do not bypass only Load Video while leaving its required-input adapter active. The current paired V3.8X2 workflows now use **Core Load Audio** and **Core Load Video -> H3 Continuum Video Adapter**. The adapter has Force Rate (default 24) and IMAGE/AUDIO outputs, but **no Enable switch or file selector**. Core owns video selection/upload/save; the established duration-preserving frame selection is reused and audio is passed unchanged. To disable Timeline Video guidance, bypass the adapter or the whole video-input group; do not bypass only Load Video while leaving its required-input adapter active.
**H3 Continuum Load Image remains**, including Enable Image. Its mode observer now delegates to inherited Core setters (Issue #23), and no longer replaces the widget collection while drawing or saving. Enable is appended after Core's widgets; Core handles the filename and native Bypass persistence. The older Audio/Video node IDs remain loadable as deprecated compatibility nodes, not as the recommended new-workflow loaders. Ten node IDs are registered including these two legacy nodes. Sampler inputs, generation parameters, Decode Cache, Finalize and existing latent/storage contracts are unchanged. **H3 Continuum Load Image remains**, including Enable Image. Its mode observer now delegates to inherited Core setters (Issue #23), and no longer replaces the widget collection while drawing or saving. Enable is appended after Core's widgets; Core handles the filename and native Bypass persistence. The older Audio/Video node IDs remain loadable as deprecated compatibility nodes, not as the recommended new-workflow loaders. Ten node IDs are registered including these two legacy nodes. Sampler inputs, generation parameters, Decode Cache, Finalize and existing latent/storage contracts are unchanged.
@@ -10,8 +10,30 @@ After updating, restart ComfyUI and reload/refresh the frontend. A workflow that
Validation details and limits: [Loader repair record](docs/LOADER_PERSISTENCE_REPAIR.md). CPU/store-model regression tests are not a substitute for the pending Windows Core 0.36.0 / frontend 1.53.6 browser tab-switch acceptance. No Release/tag or Registry publication is implied by this main-branch repair. Older loader screenshots below describe the retained legacy nodes. Validation details and limits: [Loader repair record](docs/LOADER_PERSISTENCE_REPAIR.md). CPU/store-model regression tests are not a substitute for the pending Windows Core 0.36.0 / frontend 1.53.6 browser tab-switch acceptance. No Release/tag or Registry publication is implied by this main-branch repair. Older loader screenshots below describe the retained legacy nodes.
## Timeline Video — Experimental on main
> **V3.8X2** is the product and workflow label for package `3.8.3`. It keeps the V3.8 production sampler and adds optional Reference Images 4–9 plus the built-in Decode Cache Helper. The older V3.8.0 package remains available from tag [`v3.8.0`](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/tree/v3.8.0).
The existing video IMAGE input is now displayed as **`Timeline Video Frames`** and has one appended optional control, **`Video Reference Mode`**. Existing workflows and API calls that do not contain the new setting keep the old behavior through the default **`Repeat Reference`** mode.
- **`Repeat Reference`** keeps the legacy Video Guide contract: each physical generation group receives the same bounded prefix from the source. It does not loop or stretch a short source.
- **`Follow Timeline`** uses the 24 fps source interval corresponding to each physical output group. Continuation-prefix frames are not consumed twice. If the source ends, the remaining groups continue without video reference; a final usable partial interval may receive only H3's short frame-grid padding.
This setting is independent of `Prompt Format = Timeline`. It does not add a new socket or loader, does not use the source video's audio automatically, and does not provide frame-perfect motion copying. The complete decoded IMAGE batch is still held in system memory, so this feature is not a long-input RAM optimization.
Use the existing connection:
```text
Core Load Video -> H3 Continuum Video Adapter -> Timeline Video Frames
```
The two comparison workflows are intentionally separate from the unchanged official V3.8X2 workflows:
- [Follow Timeline JSON](examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.json) | [ZIP](examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.zip)
- [Repeat Reference JSON](examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.json) | [ZIP](examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.zip)
Local acceptance passed with full CPU `1424 passed / 1 skipped / 0 failed`, browser save/reload, and Follow/Repeat `2 x 5 s` GPU runs at 704x416, 24 fps, 240 frames, and 10-second 32 kHz stereo audio. Follow changed to the later source interval for Chunk 2 while Repeat retained the shared prefix; both completed without OOM, NaN, allocation failure, or crash. Peak observed VRAM was about 15.6–15.7 GiB on the tested RTX 5060 Ti 16 GB system, so 16 GB headroom was small. This is a functional mode-separation result, not a universal speed or subjective-quality guarantee. See [Timeline Video Experimental](docs/TIMELINE_VIDEO_EXPERIMENTAL.md).
> **V3.8X2** is the product and workflow label for package `3.8.3`. It keeps the V3.8 production sampler and adds optional Reference Images 4–9 plus the built-in Decode Cache Helper. The older V3.8.0 package remains available from tag [`v3.8.0`](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/tree/v3.8.0).
<img width="1536" height="1024" alt="exec-9cfa73b4-c1d3-4416-8950-fa661b946638_2" src="https://github.com/user-attachments/assets/8df15764-6db6-4d4c-9eb5-0d2c7b0668f0" /> <img width="1536" height="1024" alt="exec-9cfa73b4-c1d3-4416-8950-fa661b946638_2" src="https://github.com/user-attachments/assets/8df15764-6db6-4d4c-9eb5-0d2c7b0668f0" />
@@ -34,7 +56,7 @@ When using many references, especially all nine, resize copies of the source ima
### Important: this is not an unconditional new-generation speedup ### Important: this is not an unconditional new-generation speedup
**Manual cache clear:** normally no action is needed. Press the Helper's "キャッシュをクリア" (Clear cache) button to discard that Helper's cache on the next Queue and run native Decode again. It does not delete immediately or queue automatically. The internal `reset_token` remains serialized for workflow/API compatibility and increments once per click (wrapping to 0 at the INT limit). Creating, loading, or cloning a node does not change it. Without JavaScript, or in API workflows, change `reset_token` through the traditional numeric input. **Manual cache clear:** normally no action is needed. Press the Helper's **Clear cache** button (Japanese UI: **キャッシュをクリア**) to discard that Helper's cache on the next Queue and run native Decode again. It does not delete immediately or queue automatically. The internal `reset_token` remains serialized for workflow/API compatibility and increments once per click (wrapping to 0 at the INT limit). Creating, loading, or cloning a node does not change it. Without JavaScript, or in API workflows, change `reset_token` through the traditional numeric input.
**The Helper saves time only when unchanged latents are decoded again. It does not accelerate Sampling.** A first MISS runs native Decode and stores the result; hashing and storage can make that run slower. Generating different latents with a new random seed each time normally produces MISS, so this feature should not be expected to speed up that usage. **The Helper saves time only when unchanged latents are decoded again. It does not accelerate Sampling.** A first MISS runs native Decode and stores the result; hashing and storage can make that run slower. Generating different latents with a new random seed each time normally produces MISS, so this feature should not be expected to speed up that usage.
@@ -50,7 +72,7 @@ The action names themselves do not enable reuse. **Each Video/Audio entry must m
Cache reuse is process-local and does not survive Python restarts. Eviction and VAE identity changes can also cause MISS. Changing mode, RAM/Disk budgets, or `reset_token` clears this implementation's Helper cache. Keeping `reset_token = 0` does not reset it on every Queue. Cache reuse is process-local and does not survive Python restarts. Eviction and VAE identity changes can also cause MISS. Changing mode, RAM/Disk budgets, or `reset_token` clears this implementation's Helper cache. Keeping `reset_token = 0` does not reset it on every Queue.
V3.8X2 ships `H3DecodeCacheHelper` inside this Continuum package. It remains a separate public node under `MiniMax H3/Continuum/Helpers`; no second custom-node addon is needed. The original nine IDs remain registered for compatibility; the loader migration adds H3ContinuumVideoAdapter. Sampler, Finalize, Assembly Plan, Core Decode, and saved Sampler widget/socket contracts are unchanged. V3.8X2 ships `H3DecodeCacheHelper` inside this Continuum package. It remains a separate public node under `MiniMax H3/Continuum/Helpers`; no second custom-node addon is needed. The package registers ten IDs, including the Video Adapter and the two deprecated loader compatibility IDs. Sampler, Finalize, Assembly Plan, Core Decode, and the pre-existing saved Sampler widget/socket prefix are unchanged; Timeline Video adds only one optional widget at the end.
Connect Sampler `video_latents` and `audio_latents` plus their native Video/Audio VAE inputs to the Helper, then connect Helper `images` and `audio` to Finalize. Keep the Sampler `assembly_plan` directly connected to Finalize. The [official V3.8X2 workflow](examples/workflows/MiniMax_H3_Continuum_V38X2.json) preserves its saved graph and settings. It contains no Core VAE Decode nodes, so returning to Core direct Decode requires adding those two Core nodes and reconnecting Finalize, or loading a prior Core-direct V3.8X workflow. Do not simply remove the Helper and expect automatic rewiring. Connect Sampler `video_latents` and `audio_latents` plus their native Video/Audio VAE inputs to the Helper, then connect Helper `images` and `audio` to Finalize. Keep the Sampler `assembly_plan` directly connected to Finalize. The [official V3.8X2 workflow](examples/workflows/MiniMax_H3_Continuum_V38X2.json) preserves its saved graph and settings. It contains no Core VAE Decode nodes, so returning to Core direct Decode requires adding those two Core nodes and reconnecting Finalize, or loading a prior Core-direct V3.8X workflow. Do not simply remove the Helper and expect automatic rewiring.
@@ -148,7 +170,7 @@ The Helper delegates cache MISS entries to native Core Video/Audio Decode. A man
## Complete V3.8 UI reference ## Complete V3.8 UI reference
This section covers every user-facing control and socket on the seven public V3.8 nodes. Names such as `Power Lora Loader (rgthree)`, Spectrum, SageAttention, Core VAE Decode, and Core Save Video belong to ComfyUI or another extension. They may be used in a supplied workflow, but they are not Continuum controls. This section covers the current ten-node V3.8X2 public surface and the Sampler controls. Names such as `Power Lora Loader (rgthree)`, Spectrum, SageAttention, Core VAE Decode, and Core Save Video belong to ComfyUI or another extension. They may be used in a supplied workflow, but they are not Continuum controls.
### H3 Continuum Sampler V3.8: required graph inputs ### H3 Continuum Sampler V3.8: required graph inputs
@@ -171,19 +193,25 @@ The Sampler returns six outputs: a list of `video_latents`, a list of `audio_lat
| `last_frame` | Optional final-image constraint for FL2VA | Can trigger a Long Terminal Merge; the final pair may become one atomic review unit | | `last_frame` | Optional final-image constraint for FL2VA | Can trigger a Long Terminal Merge; the final pair may become one atomic review unit |
| `reference_image_1`–`reference_image_3` | Ordered appearance, identity, subject, or scene references | They never become the implicit size source | | `reference_image_1`–`reference_image_3` | Ordered appearance, identity, subject, or scene references | They never become the implicit size source |
| `Reference Images (Optional)` | Additional images 4–9 bundled by `H3 Continuum Reference Images` | Empty inputs are skipped; encoding remains in the Sampler | | `Reference Images (Optional)` | Additional images 4–9 bundled by `H3 Continuum Reference Images` | Empty inputs are skipped; encoding remains in the Sampler |
| `Video Guide Frames` | A video loader's IMAGE frame batch, applied as a persistent guide | It does not carry the source video's audio; frames are interpreted at 24 fps | | `Timeline Video Frames` | The Video Adapter's 24 fps IMAGE frame batch | `Repeat Reference` keeps the legacy shared prefix; experimental `Follow Timeline` selects successive physical output intervals. Source audio is separate. |
| `Driving Audio` + `Driving Audio VAE` | Original audio timeline used as native guide conditioning | The selected source audio becomes final audio; generated audio and Audio Seam are bypassed | | `Driving Audio` + `Driving Audio VAE` | Original audio timeline used as native guide conditioning | The selected source audio becomes final audio; generated audio and Audio Seam are bypassed |
| `Reference Audio (Optional)` + `Reference Audio VAE (Optional)` | Legacy single conditioning-only audio reference | Generated audio remains final audio | | `Reference Audio (Optional)` + `Reference Audio VAE (Optional)` | Legacy single conditioning-only audio reference | Generated audio remains final audio |
| `Audio References (Optional)` | Ordered bundle from `H3 Continuum Reference Audios` | Do not connect this together with the legacy single Reference Audio path | | `Audio References (Optional)` | Ordered bundle from `H3 Continuum Reference Audios` | Do not connect this together with the legacy single Reference Audio path |
| `Still Image Guide (Optional)` | Compatibility socket inherited from the V3.7 guide contract | Experimental; not part of the V3.8X2 official workflow | | `Still Image Guide (Optional)` | Compatibility socket inherited from the V3.7 guide contract | Experimental; not part of the V3.8X2 official workflow |
#### First Image, Last Image, and Reference Images #### First Image, Last Image, and Reference Images
![First Image enabled and Last Image bypassed](docs/images/v38-manual/first-last-image-bypass.png) ![First Image enabled and Last Image bypassed](docs/images/v38-manual/first-last-image-bypass.png)
- **First Image** establishes the opening image and is the normal visual starting point for I2VA/FL2VA. It is also the only media input that can drive `Size Source = First Image`. - **First Image** establishes the opening image and is the normal visual starting point for I2VA/FL2VA. It is also the only media input that can drive `Size Source = First Image`.
- **Last Image** constrains the sequence ending. Leave it `OFF` for T2VA and ordinary I2VA. With a connected Last Image, extending or regenerating the sequence may rebuild the terminal pair. - **Last Image** constrains the sequence ending. Leave it `OFF` for T2VA and ordinary I2VA. With a connected Last Image, extending or regenerating the sequence may rebuild the terminal pair.
- **Reference Images 1–9** guide identity or appearance throughout generation. Inputs 1–3 remain on the Sampler; optional inputs 4–9 are on `H3 Continuum Reference Images`, connected to the Sampler's `Reference Images (Optional)` socket. The helper only bundles images: resizing, hashing, and VAE encoding remain in the existing Sampler path. Empty slots are skipped; Picture numbers follow active slot order after any First/Last Image. Existing three-loader templates still work unchanged. - **Reference Images 1–9** guide identity or appearance throughout generation. Inputs 1–3 remain on the Sampler; optional inputs 4–9 are on `H3 Continuum Reference Images`, connected to the Sampler's `Reference Images (Optional)` socket. The helper only bundles images: resizing, hashing, and VAE encoding remain in the existing Sampler path. Empty slots are skipped; Picture numbers follow active slot order after any First/Last Image. Existing three-loader templates still work unchanged.
#### Timeline Video Frames and Video Reference Mode
Connect `Core Load Video` to `H3 Continuum Video Adapter`, then connect Adapter `images` to `Timeline Video Frames`. The Adapter's default `Force Rate = 24` matches the input contract. Connect Adapter `audio` to `Driving Audio` only when that source audio should guide generation and become the final audio.
`Video Reference Mode` is visible only while Timeline Video Frames is connected. `Repeat Reference` is the compatibility default. `Follow Timeline` advances through the source according to the physical visible frames actually added to the output, including Review, Resume, and Terminal Merge grouping. It does not change `Seconds per Chunk`; all chunks still use the Sampler's common duration. See the [experimental contract and limitations](docs/TIMELINE_VIDEO_EXPERIMENTAL.md).
Saved workflows with the earlier direct image 4/5 inputs migrate those links into a bundle after graph loading. Migration preserves the original source links and only removes a legacy input after its new link is verified. If a source is unavailable or its destination is occupied, the legacy link is retained with a warning. Old API prompts remain accepted; supplying both legacy and bundle images for the same slot is an explicit input conflict. Saved workflows with the earlier direct image 4/5 inputs migrate those links into a bundle after graph loading. Migration preserves the original source links and only removes a legacy input after its new link is verified. If a source is unavailable or its destination is occupied, the legacy link is retained with a warning. Old API prompts remain accepted; supplying both legacy and bundle images for the same slot is an explicit input conflict.
@@ -201,7 +229,7 @@ Every `H3 Continuum Load Image` has `Enable Image`. `ON` loads through ComfyUI C
![Continuum Video and Audio loaders with Enable controls](docs/images/v38-manual/input-video-audio-bypass.png) ![Continuum Video and Audio loaders with Enable controls](docs/images/v38-manual/input-video-audio-bypass.png)
- `H3 Continuum Load Video`: `Enable Video`, `Video`, and `Force Rate`. `Force Rate = 0` uses source FPS; a positive value drops/duplicates frames to that rate while preserving nominal duration and audio. Connect `images` to `Video Guide Frames` when visual video guidance is wanted. Connect `audio` to `Driving Audio` only when the source audio should guide and become the final output. - New workflows use Core `Load Video` followed by `H3 Continuum Video Adapter`. The Adapter's positive `Force Rate` drops/duplicates frames while preserving nominal duration and passes audio through unchanged. Connect `images` to `Timeline Video Frames`; connect `audio` to `Driving Audio` only when the source audio should guide and become the final output. The legacy `H3 Continuum Load Video` remains registered for compatibility.
- `H3 Continuum Load Audio`: `Enable Audio` and the audio file. Use it for Driving Audio or Reference Audio according to the socket you connect. - `H3 Continuum Load Audio`: `Enable Audio` and the audio file. Use it for Driving Audio or Reference Audio according to the socket you connect.
- The node titled `Audio Switch` in the example image is ComfyUI Core's `If/Else Switch` (`ComfySwitchNode`); `Fast Groups Bypasser (rgthree)` belongs to the rgthree extension. Neither is a Continuum node or a requirement of Continuum. The Continuum loaders already provide their own native ON/OFF controls. - The node titled `Audio Switch` in the example image is ComfyUI Core's `If/Else Switch` (`ComfySwitchNode`); `Fast Groups Bypasser (rgthree)` belongs to the rgthree extension. Neither is a Continuum node or a requirement of Continuum. The Continuum loaders already provide their own native ON/OFF controls.
@@ -249,7 +277,7 @@ Every `H3 Continuum Load Image` has `Enable Image`. `ON` loads through ComfyUI C
![Manual Size Source with explicit Width and Height](docs/images/v38-manual/size-source-manual-640.png) ![Manual Size Source with explicit Width and Height](docs/images/v38-manual/size-source-manual-640.png)
If First Image mode is selected but no usable First Image reaches the Sampler, V3.8 uses the stored Manual Width/Height as a safe fallback and reports the dimensions in `status`. Select `Size Source = Manual` to inspect or edit those values; Width/Height are hidden in the First Image facade. Reference Images and Video Guide never become implicit size sources. The historical `Auto / Landscape / Portrait / Square` values are migration/API compatibility values, not current Main choices. If First Image mode is selected but no usable First Image reaches the Sampler, V3.8 uses the stored Manual Width/Height as a safe fallback and reports the dimensions in `status`. Select `Size Source = Manual` to inspect or edit those values; Width/Height are hidden in the First Image facade. Reference Images and Timeline Video Frames never become implicit size sources. The historical `Auto / Landscape / Portrait / Square` values are migration/API compatibility values, not current Main choices.
Size selection is built into the V3.8 Sampler. A separate megapixel, empty-latent, or image-size node is not required for the normal Continuum workflow. Size selection is built into the V3.8 Sampler. A separate megapixel, empty-latent, or image-size node is not required for the normal Continuum workflow.
@@ -271,7 +299,7 @@ For an audio-only listening diagnostic, use T2VA, `Size Source = Manual`, and a
| `Variation Nonce` | Explicit `Regenerate From = Chunk N` | `0`: automatic variation selection with interrupted-run resume; `1` or higher: a fixed variation value. Base Seed and the other generation settings stay unchanged | | `Variation Nonce` | Explicit `Regenerate From = Chunk N` | `0`: automatic variation selection with interrupted-run resume; `1` or higher: a fixed variation value. Base Seed and the other generation settings stay unchanged |
| `Run Name (Optional Override)` | Advanced open and Progress On | Stable name for resume and Render History; blank uses the Sampler's automatic identity | | `Run Name (Optional Override)` | Advanced open and Progress On | Stable name for resume and Render History; blank uses the Sampler's automatic identity |
| `Reference Image Size` | Advanced open and a Reference Image connected | `Match Output` is practical; `Max Identity` preserves more reference detail and may use more memory | | `Reference Image Size` | Advanced open and a Reference Image connected | `Match Output` is practical; `Max Identity` preserves more reference detail and may use more memory |
| `Video Guide Size` | Advanced open and Video Guide connected | `Efficient — 0.4 MP`, `Balanced — 0.6 MP`, or `Match Output` | | `Video Guide Size` | Advanced open and Timeline Video Frames connected | `Efficient — 0.4 MP`, `Balanced — 0.6 MP`, or `Match Output` |
Frontend-managed IDs, selected Take IDs, one-shot review actions, diagnostics, preview, and legacy compatibility values are deliberately not editable as ordinary node widgets. The corresponding settings or action buttons below are the supported interface. Frontend-managed IDs, selected Take IDs, one-shot review actions, diagnostics, preview, and legacy compatibility values are deliberately not editable as ordinary node widgets. The corresponding settings or action buttons below are the supported interface.
@@ -355,7 +383,7 @@ The same graph can be used as the quick Turbo path: disable Spectrum, enable exa
An experimental prefix-amplitude renormalization produced a strong improvement in one `6 × 5 s` run, but a later frozen `4 × 8 s` A/B/C gate did not reproduce the target drift and therefore applied gains of `1.0`; Standard and the experimental arm were bit-exact. The older Compatibility route did not justify replacing Standard and raised a separate boundary-audio concern. The accurate release statement is therefore: **the issue is not proven fixed, the current Standard path is unchanged, and experimental mitigation stays Default Off**. Report the first and worst affected chunk, workflow JSON, prompt, seed, model/LoRA, Steps/SIGMAS, size, Continuity, and Audio Continuity when reproducing it. An experimental prefix-amplitude renormalization produced a strong improvement in one `6 × 5 s` run, but a later frozen `4 × 8 s` A/B/C gate did not reproduce the target drift and therefore applied gains of `1.0`; Standard and the experimental arm were bit-exact. The older Compatibility route did not justify replacing Standard and raised a separate boundary-audio concern. The accurate release statement is therefore: **the issue is not proven fixed, the current Standard path is unchanged, and experimental mitigation stays Default Off**. Report the first and worst affected chunk, workflow JSON, prompt, seed, model/LoRA, Steps/SIGMAS, size, Continuity, and Audio Continuity when reproducing it.
> **Support boundary:** The original V3.8 launch surface had seven nodes; this local package also exports Reference Images and Decode Cache Helper. Some retain earlier IDs, including Finalize (`H3ContinuumAssembleSeamV35`) and Second Pass (`H3ContinuumSecondPassV35`); this does not make all older workflows compatible. Use the matching historical Release/tag for unsupported node IDs. See [V3.8 Release and Migration Policy](docs/V38_RELEASE_AND_MIGRATION.md). > **Support boundary:** The original V3.8 launch surface had seven nodes. V3.8X2 now exports ten IDs: it adds Reference Images, Decode Cache Helper, and Video Adapter while retaining the deprecated Audio/Video loader IDs for saved-workflow compatibility. Some nodes retain earlier IDs, including Finalize (`H3ContinuumAssembleSeamV35`) and Second Pass (`H3ContinuumSecondPassV35`); this does not make all older workflows compatible. Use the matching historical Release/tag for unsupported node IDs. See [V3.8 Release and Migration Policy](docs/V38_RELEASE_AND_MIGRATION.md).
The accepted 16GB GPU gates reached about `15.5-15.6 GiB` in the most complex cases. Actual use varies by GPU, driver, backend, model precision, resolution, and connected nodes; this is not a universal 16GB guarantee. The accepted 16GB GPU gates reached about `15.5-15.6 GiB` in the most complex cases. Actual use varies by GPU, driver, backend, model precision, resolution, and connected nodes; this is not a universal 16GB guarantee.
@@ -712,7 +740,7 @@ Prompt/CLIP figures measure only the conditioning subphase, not total generation
The measured Sage-only production baselines on the tested RTX 5060 Ti 16 GB / 64 GB system were 168.069 seconds for 1×5-second 576×576 T2VA and 379.765 seconds for 3×5-second 640×640 FL2VA Long Terminal Merge. These are configuration-specific baselines, not universal speed guarantees. Sampling remained the dominant cost; Continuum Assemble + Seam stayed below 1%. The measured Sage-only production baselines on the tested RTX 5060 Ti 16 GB / 64 GB system were 168.069 seconds for 1×5-second 576×576 T2VA and 379.765 seconds for 3×5-second 640×640 FL2VA Long Terminal Merge. These are configuration-specific baselines, not universal speed guarantees. Sampling remained the dominant cost; Continuum Assemble + Seam stayed below 1%.
> **V3.8.0 is the historical release baseline.** The current publication target is V3.8X2 / package 3.8.3. Historical implementation modules remain in source because V3.8X2 reuses them internally. Only the nine current public nodes are exported, including the earlier IDs retained for those nodes. Use the matching historical Release/tag for workflows requiring other IDs. Still Image Guide remains Experimental. > **V3.8.0 is the historical release baseline.** The current publication target is V3.8X2 / package 3.8.3. Historical implementation modules remain in source because V3.8X2 reuses them internally. Only the ten current public node IDs are exported, including the two deprecated loader compatibility IDs. Use the matching historical Release/tag for workflows requiring other IDs. Still Image Guide remains Experimental.
## V3.5.1 Reference Audio & Compatibility Update ## V3.5.1 Reference Audio & Compatibility Update
@@ -723,7 +751,7 @@ V3.5.1 added two focused features without changing the V3.4 Sampling, Conditioni
The Reference Audio sockets are permanently defined by Python `INPUT_TYPES`. Dynamic socket changes and the UI-only `Hidden / Show` control were removed to prevent workflow save/reload value shifts. Node IDs, backend keys, Sampling, Conditioning, standard Seed handling, and other widgets are unchanged. The Reference Audio sockets are permanently defined by Python `INPUT_TYPES`. Dynamic socket changes and the UI-only `Hidden / Show` control were removed to prevent workflow save/reload value shifts. Node IDs, backend keys, Sampling, Conditioning, standard Seed handling, and other widgets are unchanged.
In the historical V3.5.1 package, V3.4 node IDs and backend socket keys remained registered for saved-workflow compatibility. V3.4/V3.5 workflows continued to load there; V3.5.1 clarified the displayed input names without rewriting saved links. V3.8 instead uses the seven-node support boundary above. In the historical V3.5.1 package, V3.4 node IDs and backend socket keys remained registered for saved-workflow compatibility. V3.4/V3.5 workflows continued to load there; V3.5.1 clarified the displayed input names without rewriting saved links. The original V3.8 launch used a seven-node boundary; current V3.8X2 exports the ten IDs listed in the current guide above.
### Reference inputs at a glance ### Reference inputs at a glance
@@ -866,7 +894,7 @@ For the failed 3 x 5s case, First Pass and the 37T Second Pass group completed.
## V3.4 compatibility baseline ## V3.4 compatibility baseline
The historical V3.5 package retained V3.4 nodes for saved workflows; it did not replace their Node IDs, public sockets, Sampling, Conditioning, Terminal Merge, Assembly, Seam, or Run Storage behavior. Those historical workflows remain available through the corresponding Release/tag. V3.8X2 does not export IDs outside its current nine public nodes, even when their implementation modules remain in source. The historical V3.5 package retained V3.4 nodes for saved workflows; it did not replace their Node IDs, public sockets, Sampling, Conditioning, Terminal Merge, Assembly, Seam, or Run Storage behavior. Those historical workflows remain available through the corresponding Release/tag. V3.8X2 does not export IDs outside its current ten public nodes, even when their implementation modules remain in source.
![H3 Continuum V3.4 workflow overview](docs/images/v34-workflow-overview.png) ![H3 Continuum V3.4 workflow overview](docs/images/v34-workflow-overview.png)
@@ -988,7 +1016,7 @@ The public nodes focus on normal production controls. Developer diagnostics and
### Timeline paths ### Timeline paths
Experimental Timeline Video and earlier timeline-audio paths remain in historical implementation code but are not exported by V3.8. Open V3.3 workflows with their historical package; new V3.8 workflows should use `Driving Audio` and `Video Guide Frames`. The retired V3.3 Timeline Sampler and earlier timeline-audio paths remain historical and are not exported by V3.8X2. The current experimental implementation instead adds `Follow Timeline` to the existing `Timeline Video Frames` input on Sampler V3.8; it does not restore the old node ID. Open V3.3 workflows with their matching historical package.
## Installation ## Installation
@@ -1005,7 +1033,7 @@ git pull --ff-only origin main
Restart ComfyUI after the update. If the node was installed with ComfyUI Manager, use its **Update** action instead of running `git pull` manually. Do not mix Manager updates and a separate Git checkout for the same installation. Restart ComfyUI after the update. If the node was installed with ComfyUI Manager, use its **Update** action instead of running `git pull` manually. Do not mix Manager updates and a separate Git checkout for the same installation.
After the backend restart, search for `H3 Continuum Sampler V3.8`. The complete V3.8X2 search surface contains the nine nodes listed above. If they are missing, check the startup console for the `H3 Continuum 3.8.3 loaded` message and any `ComfyUI-H3-Continuum` import error. After the backend restart, search for `H3 Continuum Sampler V3.8`. The complete V3.8X2 search surface contains the ten nodes listed above. If they are missing, check the startup console for the `H3 Continuum 3.8.3 loaded` message and any `ComfyUI-H3-Continuum` import error.
Search for H3 Continuum or Continuum in ComfyUI Manager, or install manually: Search for H3 Continuum or Continuum in ComfyUI Manager, or install manually:
@@ -1315,9 +1343,9 @@ Use a 24 fps source for `Video Guide Frames`. `Load Video (Upload)` may accept f
## Current validation status ## Current validation status
The V3.8X2 public-surface suite checks the exact nine-node export, both official workflow names and matching ZIP payloads, declared external dependencies, Registry exclusions, preserved V3.8 widget/socket order, and presentation-only `Show Advanced Settings` / `Hide Advanced Settings` behavior. Registry payload hashes are recorded in `REGISTRY_MANIFEST.sha256`; source integrity is recorded in `MANIFEST.sha256`. The V3.8X2 public-surface suite checks the exact ten-node export, both unchanged official workflow names and matching ZIP payloads, the separate experimental Follow/Repeat workflows, declared external dependencies, Registry exclusions, the preserved V3.8 widget/socket prefix plus the appended mode widget, and presentation-only `Show Advanced Settings` / `Hide Advanced Settings` behavior. Registry payload hashes are recorded in `REGISTRY_MANIFEST.sha256`; source integrity is recorded in `MANIFEST.sha256`.
**Latest V3.8X2 preparation evidence:** the full CPU suite passed **1,367 passed / 1 skipped / 0 failed**. The prior dedicated GPU Functional Gate passed a `3 × 5 second` Review Each Chunk run: Q1–Q3 generated one physical group at a time, and Q4 reused all three groups with `3 reused / 0 generated`. Q3 and Q4 had identical decoded RGB and PCM SHA-256 values. The separate nine-reference Gate is recorded below. These checks confirm functional execution, prefix reuse, AV reconstruction, and the tested nine-reference path; they are not a full subjective image- or audio-quality rating. **Latest V3.8X2 preparation evidence:** after the Timeline Video integration, the full CPU suite passed **1,424 passed / 1 skipped / 0 failed**. Browser persistence and Follow/Repeat `2 × 5 second` GPU gates passed; both produced 704×416, 24 fps, 240-frame, 10-second video with 32 kHz stereo audio and no OOM, NaN, allocation failure, or crash. The prior dedicated Review Functional Gate passed a `3 × 5 second` run: Q1–Q3 generated one physical group at a time, and Q4 reused all three groups with `3 reused / 0 generated`; Q3 and Q4 had identical decoded RGB and PCM SHA-256 values. The separate nine-reference Gate is recorded below. These checks confirm functional execution, prefix reuse, AV reconstruction, mode separation, and the tested nine-reference path; they are not a full subjective image- or audio-quality rating.
### Nine-reference input gate (0.3 MP reference images) ### Nine-reference input gate (0.3 MP reference images)
+38 -9
View File
@@ -10,6 +10,28 @@
検証範囲は[Loader修正記録](docs/LOADER_PERSISTENCE_REPAIR.md)を参照してください。CPU回帰テストと実ブラウザ確認は別です。Windows Core 0.36.0 / frontend 1.53.6のタブ切替確認は未実施で、Release/tag・Registry公開も別作業です。以下の旧Loader画像は互換ノードの説明です。 検証範囲は[Loader修正記録](docs/LOADER_PERSISTENCE_REPAIR.md)を参照してください。CPU回帰テストと実ブラウザ確認は別です。Windows Core 0.36.0 / frontend 1.53.6のタブ切替確認は未実施で、Release/tag・Registry公開も別作業です。以下の旧Loader画像は互換ノードの説明です。
## Timeline Video — main上のExperimental機能
従来の動画IMAGE入力を**`Timeline Video Frames`**と表示し、末尾に省略可能な設定**`Video Reference Mode`**を1項目追加しました。新しい設定を持たない既存Workflow/APIは、初期値の**`Repeat Reference`**で従来動作を維持します。
- **`Repeat Reference`**:従来のVideo Guide契約です。各physical生成groupで、同じ先頭側の上限付き参照を使います。短い素材を自動ループ/Stretchしません。
- **`Follow Timeline`**:各physical出力groupの時間位置に対応する24fps参照区間を使います。Continuationの過去prefixを二重消費しません。素材終了後は動画参照なしで続行し、利用可能な終端区間にはH3 frame-grid整列分だけ短いpaddingを行う場合があります。
これは`Prompt Format = Timeline`とは別の設定です。新しいsocketやLoaderは増やさず、元動画のAudioも自動使用しません。参照動画はsoft conditioningであり、motionをframe単位で完全コピーする機能ではありません。また、decode済みIMAGE batch全体をsystem RAMへ保持するため、長尺入力のRAM削減機能ではありません。
接続は既存のままです。
```text
Core Load Video → H3 Continuum Video Adapter → Timeline Video Frames
```
公式V3.8X2 Workflow本体は変更せず、比較用Experimental Workflowを別名で追加しています。
- [Follow Timeline JSON](examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.json)/[ZIP](examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.zip)
- [Repeat Reference JSON](examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.json)/[ZIP](examples/workflows/MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.zip)
ローカル受入ではFull CPU `1424 passed / 1 skipped / 0 failed`、ブラウザ保存・再読込、Follow/Repeatの`2 × 5秒` GPUテストがPASSしました。両方とも704×416、24fps、240フレーム、映像10秒、32kHz stereo音声10秒で、OOM/NaN/allocation failure/crashはありません。FollowはChunk 2で後半区間へ切り替わり、Repeatは同じ先頭側参照を維持しました。RTX 5060 Ti 16 GBで観測した最大VRAMは約15.6~15.7 GiBで、16 GBの余裕は小さい条件です。これは機能・mode分離の受入結果であり、一般的な速度や主観画質の保証ではありません。詳細は[Timeline Video Experimental](docs/TIMELINE_VIDEO_EXPERIMENTAL.md)を参照してください。
> **V3.8X2**はpackage `3.8.3`の製品名・Workflow名です。V3.8 Production Samplerを維持し、任意のReference Image 4~9と内蔵Decode Cache Helperを追加しています。旧V3.8.0は[`v3.8.0`タグ](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/tree/v3.8.0)から導入できます。 > **V3.8X2**はpackage `3.8.3`の製品名・Workflow名です。V3.8 Production Samplerを維持し、任意のReference Image 4~9と内蔵Decode Cache Helperを追加しています。旧V3.8.0は[`v3.8.0`タグ](https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/tree/v3.8.0)から導入できます。
@@ -48,7 +70,7 @@ Reference Image 1~3はSamplerへ直接接続し、任意の4~9は`H3 Continu
キャッシュはPythonプロセス内の利用に限られ、再起動後は再利用しません。上限による退避・破棄やVAE識別情報の変化でもMISSになります。モード、RAM/Disk予算、`reset_token`を変更すると、この実装ではHelperキャッシュをクリアします。`reset_token = 0`を維持することは、毎回リセットする意味ではありません。 キャッシュはPythonプロセス内の利用に限られ、再起動後は再利用しません。上限による退避・破棄やVAE識別情報の変化でもMISSになります。モード、RAM/Disk予算、`reset_token`を変更すると、この実装ではHelperキャッシュをクリアします。`reset_token = 0`を維持することは、毎回リセットする意味ではありません。
`H3DecodeCacheHelper`をContinuumパッケージ内に収録し、UI上は`MiniMax H3/Continuum/Helpers`の独立公開ノードとして残します。インストール対象はContinuumだけです。公開IDはV3.8Xの7件にReference ImagesとHelperを加えた計9件です。Sampler、Finalize、Assembly Plan、Core Decode、Samplerの保存済みwidget/socket契約は変更しません。 `H3DecodeCacheHelper`をContinuumパッケージ内に収録し、UI上は`MiniMax H3/Continuum/Helpers`の独立公開ノードとして残します。インストール対象はContinuumだけです。現行パッケージはVideo Adapterと互換用旧Audio/Video Loaderを含む計10 IDを登録します。Sampler、Finalize、Assembly Plan、Core Decode、既存Sampler widget/socketの先頭契約は変更せず、Timeline Videoのmode widgetだけをoptional末尾へ追加します。
Samplerの`video_latents`/`audio_latents`と同じVideo/Audio VAEをHelperへ接続し、Helperの`images`/`audio`をFinalizeへ接続します。Samplerの`assembly_plan`はFinalizeへ直接接続したままです。[公式V3.8X2 Workflow](examples/workflows/MiniMax_H3_Continuum_V38X2.json)は保存済みgraphと設定を維持します。このgraphにはCore VAE Decodeノードがないため、Core直結へ戻すにはCore Video/Audio Decodeを追加してFinalizeへ再配線するか、旧V3.8XのCore直結graphを開いてください。Helperを単に削除しても自動では戻りません。 Samplerの`video_latents`/`audio_latents`と同じVideo/Audio VAEをHelperへ接続し、Helperの`images`/`audio`をFinalizeへ接続します。Samplerの`assembly_plan`はFinalizeへ直接接続したままです。[公式V3.8X2 Workflow](examples/workflows/MiniMax_H3_Continuum_V38X2.json)は保存済みgraphと設定を維持します。このgraphにはCore VAE Decodeノードがないため、Core直結へ戻すにはCore Video/Audio Decodeを追加してFinalizeへ再配線するか、旧V3.8XのCore直結graphを開いてください。Helperを単に削除しても自動では戻りません。
@@ -99,13 +121,14 @@ V3.8X2 Workflow:[JSON](examples/workflows/MiniMax_H3_Continuum_V38X2.json)/[
H3 Continuumは、長尺映像を最初からやり直さずに生成・確認・部分再生成・再開するProduction Samplerです。V3.8は`Main / Production`と`Advanced`の2層に整理しました。 H3 Continuumは、長尺映像を最初からやり直さずに生成・確認・部分再生成・再開するProduction Samplerです。V3.8は`Main / Production`と`Advanced`の2層に整理しました。
このローカル版で検索可能なノードは、元のV3.8の7ノードにReference ImagesとDecode Cache Helperを加えた次の9つです。 現行V3.8X2で検索可能なノードIDは次の10個です。旧Audio/Video Loaderは保存Workflow互換用で、新規WorkflowはCore LoaderとVideo Adapterを使います。
- `H3 Continuum Sampler V3.8` - `H3 Continuum Sampler V3.8`
- `H3 Continuum Finalize` - `H3 Continuum Finalize`
- `H3 Continuum Load Image` - `H3 Continuum Load Image`
- `H3 Continuum Load Audio` - `H3 Continuum Load Audio`
- `H3 Continuum Load Video` - `H3 Continuum Load Video`
- `H3 Continuum Video Adapter`
- `H3 Continuum Second Pass` - `H3 Continuum Second Pass`
- `H3 Continuum Reference Audios` - `H3 Continuum Reference Audios`
- `H3 Continuum Reference Images` - `H3 Continuum Reference Images`
@@ -119,12 +142,18 @@ V3.8X2公式Workflowの標準経路は`Sampler V3.8 -> Decode Cache Helper -> Fi
`H3 Continuum Reference Audios`は、最大3本の単独Reference Audioを1本の`Audio References (Optional)` socketへまとめます。入力は間を空けずに接続し、Promptでは接続順どおりに`<Audio 1>`、`<Audio 2>`、`<Audio 3>`を使用します。共通のCore Audio VAEで各Audioを個別encodeします。生成後の最終Audioを置換せず、Driving Audio契約も変更しません。既存Workflowは従来の単数`Reference Audio (Optional)`をそのまま使用できますが、単数経路とbundle経路は同時接続しないでください。 `H3 Continuum Reference Audios`は、最大3本の単独Reference Audioを1本の`Audio References (Optional)` socketへまとめます。入力は間を空けずに接続し、Promptでは接続順どおりに`<Audio 1>`、`<Audio 2>`、`<Audio 3>`を使用します。共通のCore Audio VAEで各Audioを個別encodeします。生成後の最終Audioを置換せず、Driving Audio契約も変更しません。既存Workflowは従来の単数`Reference Audio (Optional)`をそのまま使用できますが、単数経路とbundle経路は同時接続しないでください。
> **対応範囲:** 元のV3.8は7ノード、このローカル版はReference ImagesとDecode Cache Helperを加えた9ノードです。Finalizeの`H3ContinuumAssembleSeamV35`、Second Passの`H3ContinuumSecondPassV35`などは旧IDを維持しますが、すべての旧Workflowに互換性があるという意味ではありません。未対応のnode IDには対応するhistorical Release/tagを使用してください。詳細は[V3.8 Release/Migration Policy](docs/V38_RELEASE_AND_MIGRATION.md)を参照してください。 > **対応範囲:** 元のV3.8 launch surfaceは7ノードでした。現行V3.8X2はReference Images、Decode Cache Helper、Video Adapterを追加し、互換用旧Audio/Video Loaderを含む計10 IDです。Finalizeの`H3ContinuumAssembleSeamV35`、Second Passの`H3ContinuumSecondPassV35`などは旧IDを維持しますが、すべての旧Workflowに互換性があるという意味ではありません。未対応のnode IDには対応するhistorical Release/tagを使用してください。詳細は[V3.8 Release/Migration Policy](docs/V38_RELEASE_AND_MIGRATION.md)を参照してください。
複雑な16GB GPU受入Gateでは約`15.5~15.6 GiB`を使用しました。GPU、driver、backend、model精度、解像度、接続ノードで変動するため、すべての16GB GPUでの動作保証ではありません。 複雑な16GB GPU受入Gateでは約`15.5~15.6 GiB`を使用しました。GPU、driver、backend、model精度、解像度、接続ノードで変動するため、すべての16GB GPUでの動作保証ではありません。
Reference Imageは最大9枚です。Samplerの直入力1~3を維持し、追加4~9は`H3 Continuum Reference Images`のoptional IMAGE入力へ接続して、出力をSamplerの`Reference Images (Optional)`へ渡します。補助ノードは画像を束ねるだけで、Resize・hash・VAE Encodeは既存Sampler経路で実行します。未接続を除き1→9の順に使い、Picture番号はFirst/Last Imageの後に連番で割り当てます。空のbundleは参照なしとなり、旧0~3枚の生成・再利用契約は変わりません。 Reference Imageは最大9枚です。Samplerの直入力1~3を維持し、追加4~9は`H3 Continuum Reference Images`のoptional IMAGE入力へ接続して、出力をSamplerの`Reference Images (Optional)`へ渡します。補助ノードは画像を束ねるだけで、Resize・hash・VAE Encodeは既存Sampler経路で実行します。未接続を除き1→9の順に使い、Picture番号はFirst/Last Imageの後に連番で割り当てます。空のbundleは参照なしとなり、旧0~3枚の生成・再利用契約は変わりません。
### Timeline Video Frames/Video Reference Mode
Core `Load Video`を`H3 Continuum Video Adapter`へ接続し、Adapterの`images`をSamplerの`Timeline Video Frames`へ接続します。Adapterの初期値`Force Rate = 24`が入力契約に一致します。元動画Audioもガイド兼最終Audioとして使う場合だけ、Adapterの`audio`を`Driving Audio`へ接続します。
`Video Reference Mode`はTimeline Video Frames接続中だけ表示します。初期値`Repeat Reference`は従来互換、`Follow Timeline`は実際に出力へ追加されるphysical visible framesに合わせて参照区間を進めます。Review、Resume、Terminal Mergeでも同じphysical group契約を使います。チャンクごとの秒数を変える機能ではなく、全チャンクは従来どおりSamplerの共通`Seconds per Chunk`を使います。詳細は[Experimental仕様と制限](docs/TIMELINE_VIDEO_EXPERIMENTAL.md)を参照してください。
以前の4・5直入力を含む保存Workflowは、グラフ読み込み後に元の画像接続をbundleへ移します。新接続を検証した後だけ旧接続口を削除し、接続元が不明・移行先が使用中の場合は警告して元の線を残します。旧API入力も維持しますが、同じ番号の画像を旧直入力とbundleの両方から指定すると入力競合として明示します。既存の3枚loaderテンプレートは上書きせず、そのまま利用できます。 以前の4・5直入力を含む保存Workflowは、グラフ読み込み後に元の画像接続をbundleへ移します。新接続を検証した後だけ旧接続口を削除し、接続元が不明・移行先が使用中の場合は警告して元の線を残します。旧API入力も維持しますが、同じ番号の画像を旧直入力とbundleの両方から指定すると入力競合として明示します。既存の3枚loaderテンプレートは上書きせず、そのまま利用できます。
### サイズの決め方を直接選択 ### サイズの決め方を直接選択
@@ -141,7 +170,7 @@ Mainの`Size Source`は次の2択です。
- `First Image`:接続したFirst Imageの縦横比を維持し、Presetの目標面積から最終Width/Heightを算出します。 - `First Image`:接続したFirst Imageの縦横比を維持し、Presetの目標面積から最終Width/Heightを算出します。
- `Manual`:Width/Heightを32 pixel単位で直接指定します。この場合、Presetは適用しません。 - `Manual`:Width/Heightを32 pixel単位で直接指定します。この場合、Presetは適用しません。
MainのWidth/Heightは`Manual`のときだけ表示・編集できます。`First Image`ではResolutionを表示し、画像の縦横比からサイズを決めます。有効なFirst ImageがSamplerへ届かない場合は、保持しているManual Width/Heightへフォールバックし、その寸法をStatusへ表示します。手動寸法を確認・変更する場合は`Size Source = Manual`を選びます。Reference ImageやVideo Guideからサイズを暗黙取得することはありません。 MainのWidth/Heightは`Manual`のときだけ表示・編集できます。`First Image`ではResolutionを表示し、画像の縦横比からサイズを決めます。有効なFirst ImageがSamplerへ届かない場合は、保持しているManual Width/Heightへフォールバックし、その寸法をStatusへ表示します。手動寸法を確認・変更する場合は`Size Source = Manual`を選びます。Reference ImageやTimeline Video Framesからサイズを暗黙取得することはありません。
旧V3.8の`Auto / Landscape / Portrait / Square`はWorkflow/API互換用として内部で受け付けます。FrontendはAutoをFirst Imageへ、3つの固定Aspectを同じ解決結果のManual Width/Heightへ移行します。 旧V3.8の`Auto / Landscape / Portrait / Square`はWorkflow/API互換用として内部で受け付けます。FrontendはAutoをFirst Imageへ、3つの固定Aspectを同じ解決結果のManual Width/Heightへ移行します。
@@ -303,7 +332,7 @@ Prompt/CLIPの数値はconditioning区間だけで、総生成時間ではあり
RTX 5060 Ti 16 GB/RAM 64 GBの検証環境で測定したSage-only Production baselineは、576×576 T2VA 1×5秒が168.069秒、640×640 FL2VA Long Terminal Merge 3×5秒が379.765秒です。環境・設定固有の測定値であり、すべての環境に対する速度保証ではありません。Samplingが最大コストで、Continuum Assemble + Seamは1%未満でした。 RTX 5060 Ti 16 GB/RAM 64 GBの検証環境で測定したSage-only Production baselineは、576×576 T2VA 1×5秒が168.069秒、640×640 FL2VA Long Terminal Merge 3×5秒が379.765秒です。環境・設定固有の測定値であり、すべての環境に対する速度保証ではありません。Samplingが最大コストで、Continuum Assemble + Seamは1%未満でした。
**V3.8.0はhistorical release baselineです。現在の公開対象はV3.8X2/package 3.8.3です。** V3.8X2が内部利用する旧module/classはsourceへ維持します。exportするのは現在の公開9ノードだけで、その一部は旧IDを維持しています。それ以外のIDを必要とする旧保存Workflowは、対応するhistorical Release/tagを使用してください。Still Image Guideは引き続きExperimentalです。 **V3.8.0はhistorical release baselineです。現在の公開対象はV3.8X2/package 3.8.3です。** V3.8X2が内部利用する旧module/classはsourceへ維持します。exportするのは現在の公開10ノードだけで、その一部は互換用の旧IDです。それ以外のIDを必要とする旧保存Workflowは、対応するhistorical Release/tagを使用してください。Still Image Guideは引き続きExperimentalです。
## V3.5.1 Reference Audio/互換性更新 ## V3.5.1 Reference Audio/互換性更新
@@ -436,7 +465,7 @@ Registry配布予定の対象もこのJSONとZIPです。依存なしのテン
## V3.4互換 ## V3.4互換
V3.4の実装moduleは内部継承とhistorical testのためsourceへ保持しています。V3.8X2からexportするのは現在の公開9ノードに含まれるIDだけです。それ以外のIDを必要とするV3.4保存Workflowは、対応するhistorical packageを使用してください。 V3.4の実装moduleは内部継承とhistorical testのためsourceへ保持しています。V3.8X2からexportするのは現在の公開10ノードに含まれるIDだけです。それ以外のIDを必要とするV3.4保存Workflowは、対応するhistorical packageを使用してください。
## V3.4.0 Stable ## V3.4.0 Stable
@@ -552,15 +581,15 @@ V3.8はComfyUI 0.34.2で検証しています。以下の歴史的な受入記
ZIPを展開して、`ComfyUI-H3-Continuum`フォルダーを`ComfyUI/custom_nodes/`へ置き、ComfyUIを再起動します。 ZIPを展開して、`ComfyUI-H3-Continuum`フォルダーを`ComfyUI/custom_nodes/`へ置き、ComfyUIを再起動します。
再起動後は`H3 Continuum Sampler V3.8`を検索してください。V3.8X2の検索面は上記9ノードです。 再起動後は`H3 Continuum Sampler V3.8`を検索してください。V3.8X2の検索面は上記10ノードです。
旧`ComfyUI-H3-Continuum-Join`が残っている場合は同時ロードを避けるため削除または退避してください。同梱`install_windows.bat`は旧名・新名の既存フォルダーを日時付きでバックアップします。 旧`ComfyUI-H3-Continuum-Join`が残っている場合は同時ロードを避けるため削除または退避してください。同梱`install_windows.bat`は旧名・新名の既存フォルダーを日時付きでバックアップします。
## 検査 ## 検査
現在のV3.8X2 Public Surface suiteは、正確な9 ID export、公式Workflow 2名称と各ZIP内JSONの同一性、外部依存、Registry除外、既存V3.8 widget/socket順、`Show Advanced Settings`/`Hide Advanced Settings`による表示切替を確認します。Registry配布対象のハッシュは`REGISTRY_MANIFEST.sha256`、source側の整合性は`MANIFEST.sha256`で管理します。 現在のV3.8X2 Public Surface suiteは、正確な10 ID export、変更していない公式Workflow 2名称と各ZIP内JSONの同一性、別名のExperimental Follow/Repeat Workflow、外部依存、Registry除外、既存V3.8 widget/socket先頭契約と末尾追加mode、`Show Advanced Settings`/`Hide Advanced Settings`による表示切替を確認します。Registry配布対象のハッシュは`REGISTRY_MANIFEST.sha256`、source側の整合性は`MANIFEST.sha256`で管理します。
**最新のV3.8X2準備証拠:** Full CPU suiteは**1,367 passed / 1 skipped / 0 failed**です。従来の最終GPU Functional Gateでは`3 × 5秒`のReview Each Chunkを実行し、Q1~Q3はphysical groupを1つずつ生成、Q4は`3 reused / 0 generated`で3 groupすべてを再利用しました。Q3とQ4の復号後RGBおよびPCM SHA-256は一致しています。9枚Reference Imageの別Gateは下記に記録しています。これは実行、prefix再利用、AV再構築、テストした9枚経路の確認であり、画像・音声の総合的な主観品質評価ではありません。 **最新のV3.8X2準備証拠:** Timeline Video統合後のFull CPU suiteは**1,424 passed / 1 skipped / 0 failed**です。ブラウザ保存・再読込とFollow/Repeatの`2 × 5秒` GPU GateもPASSし、両方とも704×416、24fps、240フレーム、映像10秒、32kHz stereo音声10秒で完走しました。OOM/NaN/allocation failure/crashはありません。従来のReview Functional Gateでは`3 × 5秒`を実行し、Q1~Q3はphysical groupを1つずつ生成、Q4は`3 reused / 0 generated`で全3 groupを再利用し、Q3とQ4の復号後RGB/PCM SHA-256が一致しました。9枚Reference Imageの別Gateは下記に記録しています。これは機能実行、prefix再利用、AV再構築、mode分離、テストした9枚経路の確認であり、画像・音声の総合的な主観品質評価ではありません。
### 9枚Reference Image入力Gate(0.3 MP参照画像) ### 9枚Reference Image入力Gate(0.3 MP参照画像)
+12 -9
View File
@@ -1,7 +1,7 @@
dca65fb1393e67f59c6f81f9411e3e745eb8f1097e9a2adf7ffc6f4e3901ca9f __init__.py dca65fb1393e67f59c6f81f9411e3e745eb8f1097e9a2adf7ffc6f4e3901ca9f __init__.py
a9fd5a4c5624e17148656abed0ba41a72c72cac73e7f6f36d2816ad92962f867 .comfyignore bb6f7d1227f7c0cc4e134ccc0e037c8c4fe4af63d8c990ac00b7af2279695871 .comfyignore
3e7a1ec3380be2a222c8a3e9b8a6bb184d3a760478957f48186438f15235bdba branch_provenance.py 3e7a1ec3380be2a222c8a3e9b8a6bb184d3a760478957f48186438f15235bdba branch_provenance.py
d70402478f19f6487eed08957a048a8c248a9291be8e62c0e7ade69da82fd393 CHANGELOG.md ad340adc9506ab069a83f586d954fbd19f44df4e18a5346d5a8923a489d3cb85 CHANGELOG.md
c1ed3a0342730450958fb62fd21854a876991f3a4aefb24e5383b0c73cd2312d compatibility.py c1ed3a0342730450958fb62fd21854a876991f3a4aefb24e5383b0c73cd2312d compatibility.py
4d62508f900af9f93ec320afb900c27cc9a300d1e0c0a94714258498c5ede5fc conditioning.py 4d62508f900af9f93ec320afb900c27cc9a300d1e0c0a94714258498c5ede5fc conditioning.py
294251c40b1be2e296c5e5f627b68091d087bdb39179e3a0c3cebd2ee8a344f5 constants.py 294251c40b1be2e296c5e5f627b68091d087bdb39179e3a0c3cebd2ee8a344f5 constants.py
@@ -35,8 +35,8 @@ e2b72149b0433e95c13659cc9e742fe38523e67d4f4a2fdb08aed1943fecf653 nodes.py
4852ad216f85c22f008abe18464eab2c913bd88baa7db35764df2de9cf401d6d PACKAGE_VALIDATION.txt 4852ad216f85c22f008abe18464eab2c913bd88baa7db35764df2de9cf401d6d PACKAGE_VALIDATION.txt
b470819f16fc179654dc1600721a63f0a5a14e777f4c004d574b499da6886ba5 public_api.py b470819f16fc179654dc1600721a63f0a5a14e777f4c004d574b499da6886ba5 public_api.py
15a3925218e0c5f968795d7d11e4783c3c6868917447274bc700328851b3fbfb pyproject.toml 15a3925218e0c5f968795d7d11e4783c3c6868917447274bc700328851b3fbfb pyproject.toml
4a0507ce6183e099265d5a34a38e43da89515af8eb680f5a48abc8aecc724d63 README_JA.md 6f108a67967dd686011bf031b8176373f6ab80635f7e01e3e935e8585b804fc9 README_JA.md
9d4eefcc97e74d4c1c8ab1be612b50eb99a8a4f8db7c2134ca85a9e67bf2568a README.md 77cdaadd32f9d722f6b05abc196cbdb9dded32387493664620a30fa405429743 README.md
44127adcfa30ee84bff9cd4b2d0c695355209550e76982d374122008728aa5bc reference_audio.py 44127adcfa30ee84bff9cd4b2d0c695355209550e76982d374122008728aa5bc reference_audio.py
5cb9bfd6810d3a93a637451980a4cc5c1436e5b84cfd1c7a111cd192e876351b reference_video.py 5cb9bfd6810d3a93a637451980a4cc5c1436e5b84cfd1c7a111cd192e876351b reference_video.py
c7b14ed6f8cfc2a1f0a21db4fe56fba75a93ffcb0be0c22bbd2e691c45cd0e2d reference.py c7b14ed6f8cfc2a1f0a21db4fe56fba75a93ffcb0be0c22bbd2e691c45cd0e2d reference.py
@@ -49,7 +49,7 @@ c7b14ed6f8cfc2a1f0a21db4fe56fba75a93ffcb0be0c22bbd2e691c45cd0e2d reference.py
f218e1a93c42467c6057c5c1d04d57cc82f6331b97ec9257e4e2dfb9c831876a timeline_video.py f218e1a93c42467c6057c5c1d04d57cc82f6331b97ec9257e4e2dfb9c831876a timeline_video.py
0bc369ee428b8a5bc8e370a81fbab427262a1fdb7805a78c442c91698d65efba tools/__init__.py 0bc369ee428b8a5bc8e370a81fbab427262a1fdb7805a78c442c91698d65efba tools/__init__.py
e5a84569953ee63b9625f4f2e3195ec368204fddcfc746564cfbd0e8449bb8e9 tools/p20_file_backed_tensor_poc.py e5a84569953ee63b9625f4f2e3195ec368204fddcfc746564cfbd0e8449bb8e9 tools/p20_file_backed_tensor_poc.py
087f4f727e9f76bb2b0061e26e0b49065370aa352a1ebebc7f253774b3be22ed tools/verify_runtime.py deeae570772c1a32dbb356aa30c8ef473b9334a0a36b71c0cccba12203f76415 tools/verify_runtime.py
ffd61f3f31c9b6430e1169007ee906c8f5000eba63ee44e2976ad4f5f1112671 v2/__init__.py ffd61f3f31c9b6430e1169007ee906c8f5000eba63ee44e2976ad4f5f1112671 v2/__init__.py
2003902c87bba09cfa143f5a9e4248f3a78c0fb544449a75fd40ba2cacb35655 v2/context_diagnostics.py 2003902c87bba09cfa143f5a9e4248f3a78c0fb544449a75fd40ba2cacb35655 v2/context_diagnostics.py
cde27db19b677e318cc4dfb4235be7a11429c19945f4083052a91fe117e74c82 v2/decoder.py cde27db19b677e318cc4dfb4235be7a11429c19945f4083052a91fe117e74c82 v2/decoder.py
@@ -62,7 +62,7 @@ dc2b901789e4b024dbcbc4024c63eb8cbc564d77003ce8c35c0584851ba6dfbd v2/prompts.py
2fd811794c4cc0f2ce050250b80484876454926b9f20fce6af3a97d728e5acdf v2/seam_guard.py 2fd811794c4cc0f2ce050250b80484876454926b9f20fce6af3a97d728e5acdf v2/seam_guard.py
be0d8585fa57e0ccd6d02577bbc7299629e5bfe20e186747f571fccc98b67d7e v2/seam_types.py be0d8585fa57e0ccd6d02577bbc7299629e5bfe20e186747f571fccc98b67d7e v2/seam_types.py
12cfcc0c71ca087fe39a5ce2922959a62b2bb311bfe664f05c4480affce03708 v2/seeds.py 12cfcc0c71ca087fe39a5ce2922959a62b2bb311bfe664f05c4480affce03708 v2/seeds.py
ff23b9f49b05b1e000110ab4eb211942f0aa0d376f35694e80c5df25047a77c7 v2/sequence.py fe7df9efea15b157c158046f64328b3dce3cfbd35cf180573ce4ec771ddcea8c v2/sequence.py
13213e7709fd02bb810af3c18e4cd7d010f8cdfb121a4b99866a9f2333331089 v2/session_io.py 13213e7709fd02bb810af3c18e4cd7d010f8cdfb121a4b99866a9f2333331089 v2/session_io.py
3ea23dbce6fcd217b6e81235855479b9b16fed7a21693cf156f9014555c4fe13 v2/session.py 3ea23dbce6fcd217b6e81235855479b9b16fed7a21693cf156f9014555c4fe13 v2/session.py
78c830cdf49034c6e3b2cd3da5914d5cfc5638eb63ec6dbd1bab13c50d6d4caf v3/__init__.py 78c830cdf49034c6e3b2cd3da5914d5cfc5638eb63ec6dbd1bab13c50d6d4caf v3/__init__.py
@@ -70,7 +70,7 @@ ea2922567df90436dd6fa21696833a49dd0cbfeee142a5fedc3a7c303824d473 v3/adaptive_co
b3dc61503d35d9ef0a0dc7fed1c9b2fdc00fcb2072d4e39f64ae0826f9ec7dca v3/assembly_v35.py b3dc61503d35d9ef0a0dc7fed1c9b2fdc00fcb2072d4e39f64ae0826f9ec7dca v3/assembly_v35.py
7815230e0a9540bc2edeaf818e940b7a1fdd817fd217ba8a4f9cb21aaf6a308e v3/assembly.py 7815230e0a9540bc2edeaf818e940b7a1fdd817fd217ba8a4f9cb21aaf6a308e v3/assembly.py
3f62fdb35e804f42d7cbf0828b4e87e69a2b99246c3f7467fe4fa698f1d803c9 v3/conditioning_bridge_nodes.py 3f62fdb35e804f42d7cbf0828b4e87e69a2b99246c3f7467fe4fa698f1d803c9 v3/conditioning_bridge_nodes.py
aa044669259df440a384b5e839f75fa9665790da0c210a427375fdd5c76af779 v3/driving_nodes.py 82238eff890574fcf977dc3d2bc092287f63710a47bfc0f6692c272814dbd295 v3/driving_nodes.py
147a1d3fc5e09b323a4c3d13d30e432410389bae52b58330462375b9f307b791 v3/easy_audio_nodes.py 147a1d3fc5e09b323a4c3d13d30e432410389bae52b58330462375b9f307b791 v3/easy_audio_nodes.py
64b3e54b42e5156642f9957ccc374fd190d9fc7145637984e702534fb1bf80b4 v3/easy_image_nodes.py 64b3e54b42e5156642f9957ccc374fd190d9fc7145637984e702534fb1bf80b4 v3/easy_image_nodes.py
ac859cb1a6e23b8ceae300d818c24fe1346b649a0d663017a9b3ac64aa96fca3 v3/easy_nodes.py ac859cb1a6e23b8ceae300d818c24fe1346b649a0d663017a9b3ac64aa96fca3 v3/easy_nodes.py
@@ -114,8 +114,11 @@ fefafd4550511d80bd85ade46a0476f0cf3de13ddb94a908dff6b175f2a0e199 web/easy_bypas
9db0aa94f634578a962d5db49e5c70356040a59df968e3a2e531b327a623081f web/easy_load_image.js 9db0aa94f634578a962d5db49e5c70356040a59df968e3a2e531b327a623081f web/easy_load_image.js
35090816ba7470a7255eb85d003b7e79f8a0065982dacf98a18f41c8ebff655d web/easy_load_video.js 35090816ba7470a7255eb85d003b7e79f8a0065982dacf98a18f41c8ebff655d web/easy_load_video.js
0c68d13a1f8beb7431a32b471f8b0612c42fcc4bf72dd97611a025d158008662 web/easy_v38.js 0c68d13a1f8beb7431a32b471f8b0612c42fcc4bf72dd97611a025d158008662 web/easy_v38.js
ec591d1b8f7587594397b995b9584809435f10fa3de6bf8f8df204c34809aa43 web/project_id.js bb7e29359e1fe26fd70d22f6bc70eb8c8858f68ab954ba39ad3670f2cdeeae7d web/project_id.js
ecdcd27528f6f1e80f758c608cbc42f747163ea4c39abd77576b91c627423d12 web/reference_audio_ui.js ae59273f163a116ece418957c557c4fd8c3285e3c60814679b9361d2e6e6fa02 web/reference_audio_ui.js
46c9f75c9462bf26300332448f4431cca801f38964ca0a49bc2c50a88a7cc40f web/reference_image_ui.js 46c9f75c9462bf26300332448f4431cca801f38964ca0a49bc2c50a88a7cc40f web/reference_image_ui.js
26b13614ba0a5fd8b192a106c9503b2f4b6fe7f9268a0fcb0f544979a49cab34 v3/video_adapter_nodes.py 26b13614ba0a5fd8b192a106c9503b2f4b6fe7f9268a0fcb0f544979a49cab34 v3/video_adapter_nodes.py
77c225d1e49e06d00b04acf811177c9a4317e792d4bfe92df539de99420c3563 docs/LOADER_PERSISTENCE_REPAIR.md 77c225d1e49e06d00b04acf811177c9a4317e792d4bfe92df539de99420c3563 docs/LOADER_PERSISTENCE_REPAIR.md
b935ad3bb729948bb3890ce967b0ecbef737a5e51ddfc47a0060fc20aa78dcda video_reference_modes.py
bf41bf08dd0504eb753fa5781e6c2b8c1285758ad4bb7309975bb9231d5f8e41 web/video_reference_mode_ui.js
4c33d223a77c78e5c74946aa43f537031fd5aa415388d46fed975f9b07d53c3e docs/TIMELINE_VIDEO_EXPERIMENTAL.md
+6
View File
@@ -1874,3 +1874,9 @@ D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-rollback-after-00038-fail-20260821_
- Updated only that expected fingerprint to the user-approved migrated workflow (1B2212B7511FAC0B63DFA6E9F006CD8644C72D08C6945AB0E4ED0E722C97A730); all Decode Cache route and output assertions remain. No runtime or workflow bytes changed in this follow-up. - Updated only that expected fingerprint to the user-approved migrated workflow (1B2212B7511FAC0B63DFA6E9F006CD8644C72D08C6945AB0E4ED0E722C97A730); all Decode Cache route and output assertions remain. No runtime or workflow bytes changed in this follow-up.
- Full isolated CPU/Node suite: 1377 passed, 3 skipped, zero failures/errors. Actions run 35558988542. Source was snapshotted with `tools/snapshot.ps1` before edits; source/Registry manifest hashes verified. - Full isolated CPU/Node suite: 1377 passed, 3 skipped, zero failures/errors. Actions run 35558988542. Source was snapshotted with `tools/snapshot.ps1` before edits; source/Registry manifest hashes verified.
- Windows deployment, live-browser checks and GPU generation were not performed; fixed H3情報チェック handoff remains pending. No Release/tag/Registry publication. - Windows deployment, live-browser checks and GPU generation were not performed; fixed H3情報チェック handoff remains pending. No Release/tag/Registry publication.
## Timeline Video Experimental local acceptance and publication preparation (2026-09-22)
- Applied the approved Timeline Video implementation to ComfyUI_W from baseline `085943ad...`. Added `Repeat Reference` compatibility default and opt-in `Follow Timeline` over the existing 24fps IMAGE input, plus slice-specific cache identity and separate comparison workflows. Official V3.8X2 workflow bytes were not changed.
- Full CPU suite passed `1424 passed / 1 skipped / 0 failed`. Browser save/reload retained Follow mode and the Timeline Video input. Follow and Repeat `2 x 5 s` GPU runs both produced 704x416/24fps/240-frame/10-second video with 32kHz stereo audio; no OOM, NaN, allocation failure, or crash. Chunk-2 A/B SSIM `0.7246` confirmed mode separation after nearly identical Chunk 1.
- README/README_JA and the experimental contract were updated with wiring, behavior, limits, measured 16GB headroom, and the corrected ten-node V3.8X2 public-surface count. No Release/tag or Registry publish is included in this main commit/push task.
+83
View File
@@ -0,0 +1,83 @@
# Unified Timeline Video — Experimental 1
## 状態
`main`向けExperimental機能です。実装基準ソース: `085943ad9cb8023a2042d4d731c83e638361c06e`。
ローカル受入はPASSです。Full CPU suiteは`1424 passed / 1 skipped / 0 failed`、実ComfyUIのブラウザ保存・再読込、Follow/RepeatのGPU生成、A/B mode分離、Manifest整合を確認しました。両GPU runは704×416、24fps、240フレーム、映像10秒、32kHz stereo音声10秒で完走し、OOM/NaN/allocation failure/crashはありません。Production標準への昇格やRelease/tag/Registry公開を意味せず、機能名どおりExperimentalです。
## 接続と設定
```
Core Load Video
↓ VIDEO
H3 Continuum Video Adapter (Force Rate = 24)
↓ images / IMAGE
H3 Continuum Sampler V3.8
└ Timeline Video Frames [内部名 reference_video_1 を維持]
```
Video AdapterとCore Load Videoは既存のままです。AudioはCore Load Audioを利用し、この機能が動画素材の音声を勝手に追加することはありません。
動画入力を有効にすると、Samplerに **Video Reference Mode** が表示されます。
この設定は **Prompt Format = Timelineとは別**です。設定を増やすのはこの1項目のみで、新しい動画ソケット、Loader、Enableは追加しません。
### Follow Timeline
生成全体の時間位置に対応する参照区間を渡します。20秒素材を2×10秒で使うと、前半グループは先頭側、後半グループはその続きです。
厳密な境界はH3の17k+5フレーム整列と既存の出力計画に従います。例えば最初の10秒グループは243フレームとなり、次の参照開始は243/24=10.125秒です。単純な10.000秒への巻戻しではありません。
対象は **実際に新しく出力へ追加される自然長** です。Continuationの過去フレームprefixは二重に消費しません。
Terminal Mergeでは、統合された1回のSamplingが担当する可視区間をまとめて使用し、Merge自体やSeedの決定を変更しません。
Resume/Reviewでは保存済みのnet_framesを積み上げる既存カウンタを使い、Queueのたびに参照開始位置を0へ戻しません。
素材が終わったら、以後は動画参照なしで続行します。終端をまたぐ場合は残りを使い、残りが5フレーム未満ならそのグループの動画参照をスキップして診断を表示します。
自動ループ、長い静止画保持、Stretchは行いません。通常のH3整列に必要な最大16フレームの短い終端補完だけは残します。
最後のグループを「今回要求した総尺」で参照クロップしない理由: 総Chunksを増やしたとき、保存済みグループの本来の参照入力が変わってしまうのを防ぐためです。実際の最終映像の尺調整は既存Assemblyのままです。
### Repeat Reference
従来のVideo Guide Frames処理をそのまま使います。毎グループで同じ先頭側の、1チャンク上限内の素材を参照します。
短い素材を必要尺まで何回もコピーする「ループ再生」ではありません。
旧ワークフローや旧APIに新設定がない場合はこのモードになります。短い/長いという素材長だけでモードを自動変更しません。
### できないこと/保証しないこと
- 元動画の動き、画素、人物、カメラを完全コピーする制御ではありません。参照映像は外観にも影響します。
- 正しい区間を渡すことと、生成結果が意図した動きに一致することは別の評価項目です。
- 単独15秒生成がstaticになるという報告やIssue #13の長尺driftの修正ではありません。
- 長い入力のRAM削減機能ではありません。Followは全24fpsフレームの正規化済みCPUコピーを保持します。
- 真のVFRタイムスタンプ追従や補間は追加していません。入力IMAGE列は24fpsという既存契約です。
## 内部構成
- `video_reference_modes.py`: Follow source、物理可視区間選択、Qwen/VAE共通の既存エンコーダ呼出し、slice identity。
- `v3/driving_nodes.py`: V38へ省略可能なscalar widgetを末尾追加。V34入力境界でモードを消費し、Repeatは元の関数と元の引数を保ちます。
- `v2/sequence.py`: 新sourceは既存の`reference_video_source`引数で運びます。旧`timeline_video_source`は使わず、Terminal Merge無効化を回避します。
- 物理グループごとに別のprompt cache辞書を作り、同じ文章でも前の区間のconditioningを再利用しません。
- VAE cacheの`source.combined_hash`にslice identityを入れます。共通キャッシュの実装は変更しません。
- Run identityはFollow/Repeat、素材内容、サイズ、chunk_seconds、前処理・終端方針を区別します。Chunks総数は含めません。
- 新規Followグループのplanに`_h3_continuum_video_reference_v1`を追加し、実参照範囲を記録します。Terminalの論理2区間には共通の物理slice情報が付きます。
- UIは既存のネイティブwidgetを保存し、独自配列入替えを追加しません。古いワークフローの読み込み時は新モードをRepeatへ復元します。
## 配布ワークフロー
適用ツールは元の公式JSONを変更せず、次の比較用コピーを生成します。各JSONと同じ内容のZIPも作成します。
- `MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Follow.json`
- `MiniMax_H3_Continuum_V38X2_Timeline_Experimental_Repeat.json`
変更はSamplerの新設定と動画入力の表示名のみです。元の画像、プロンプト、Seed、Run名、バイパス状態、外部ノード設定・接続は維持します。
元の動画入力グループがBypassなら、そのままです。試験時にグループ全体を有効化してください。Core Load Videoだけを無効にしてAdapterを有効のままにしないでください。
## 受入結果と残る境界
1. Full CPU suite: `1424 passed / 1 skipped / 0 failed`。
2. ブラウザ: Followを保存し、再読込後も`Video Reference Mode`と`Timeline Video Frames`を保持。
3. Follow GPU: Chunk 1とChunk 2で別の参照区間を使用し、区間ごとに別conditioning/VAE cache identityを確認。
4. Repeat GPU: 2 physical groupsで従来のpersistent参照を維持し、VAE conditioningはrun内で1回。
5. A/B: Chunk 1はほぼ一致し、Chunk 2から明確に分岐。2 modeが同じconditioning経路へ潰れていないことを確認。
6. Integrity: 出力仕様、AV duration、Manifest/Registry Manifest、OOM/NaN/crashなしを確認。
未検証の全モデル・解像度・長尺・source exhaustion・Resume/Retry組合せを一括保証しません。参照区間が正しく選ばれることと、生成映像が意図したmotionを忠実に再現することは別の評価です。Production標準への昇格、Release/tag、Registry公開は別判断です。
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
+9 -2
View File
@@ -76,9 +76,16 @@ def test_public_helper_and_v38_socket_are_the_only_surface_additions():
) )
assert H3ContinuumReferenceAudios.RETURN_TYPES == (REFERENCE_AUDIOS_TYPE,) assert H3ContinuumReferenceAudios.RETURN_TYPES == (REFERENCE_AUDIOS_TYPE,)
assert "audio_references" not in v37["optional"] assert "audio_references" not in v37["optional"]
assert list(v38["optional"])[:-2] == list(v37["optional"]) assert list(v38["optional"])[:-3] == list(v37["optional"])
assert list(v38["optional"])[-2:] == ["audio_references", "image_references"] assert list(v38["optional"])[-3:] == [
"audio_references",
"image_references",
"video_reference_mode",
]
assert v38["optional"]["audio_references"][0] == REFERENCE_AUDIOS_TYPE assert v38["optional"]["audio_references"][0] == REFERENCE_AUDIOS_TYPE
mode_schema = v38["optional"]["video_reference_mode"]
assert mode_schema[0] == ("Follow Timeline", "Repeat Reference")
assert mode_schema[1]["default"] == "Repeat Reference"
def test_helper_preserves_order_and_requires_contiguous_core_numbering(): def test_helper_preserves_order_and_requires_contiguous_core_numbering():
+5
View File
@@ -12,6 +12,9 @@ from ComfyUI_H3_Continuum_Join import nodes as root_nodes
from ComfyUI_H3_Continuum_Join.v3 import driving_nodes from ComfyUI_H3_Continuum_Join.v3 import driving_nodes
from ComfyUI_H3_Continuum_Join.v3 import easy_nodes from ComfyUI_H3_Continuum_Join.v3 import easy_nodes
from ComfyUI_H3_Continuum_Join.v3 import second_pass_nodes from ComfyUI_H3_Continuum_Join.v3 import second_pass_nodes
from ComfyUI_H3_Continuum_Join.tools.verify_runtime import (
EXPECTED_PUBLIC_NODE_DISPLAY_NAMES,
)
ROOT = Path(__file__).resolve().parents[1] ROOT = Path(__file__).resolve().parents[1]
@@ -32,6 +35,7 @@ PUBLIC_DISPLAY_NAMES = {
def test_public_export_keeps_legacy_loaders_and_adds_video_adapter(): def test_public_export_keeps_legacy_loaders_and_adds_video_adapter():
assert set(root_nodes.NODE_CLASS_MAPPINGS) == set(PUBLIC_DISPLAY_NAMES) assert set(root_nodes.NODE_CLASS_MAPPINGS) == set(PUBLIC_DISPLAY_NAMES)
assert root_nodes.NODE_DISPLAY_NAME_MAPPINGS == PUBLIC_DISPLAY_NAMES assert root_nodes.NODE_DISPLAY_NAME_MAPPINGS == PUBLIC_DISPLAY_NAMES
assert EXPECTED_PUBLIC_NODE_DISPLAY_NAMES == PUBLIC_DISPLAY_NAMES
def test_labs_and_unreleased_easy_implementations_remain_internal(): def test_labs_and_unreleased_easy_implementations_remain_internal():
@@ -110,6 +114,7 @@ def test_v38_schema_and_serialized_widget_order_are_unchanged():
"guide", "guide",
"audio_references", "audio_references",
"image_references", "image_references",
"video_reference_mode",
] ]
+9 -2
View File
@@ -87,11 +87,18 @@ def test_current_v38_public_schema_is_preserved_exactly():
) )
) )
assert tuple(v38["required"]) == tuple(expected_required) assert tuple(v38["required"]) == tuple(expected_required)
assert list(v38["optional"])[:-2] == list(v37["optional"]) assert list(v38["optional"])[:-3] == list(v37["optional"])
assert list(v38["optional"])[-2:] == ["audio_references", "image_references"] assert list(v38["optional"])[-3:] == [
"audio_references",
"image_references",
"video_reference_mode",
]
assert v38["optional"]["audio_references"][0] == ( assert v38["optional"]["audio_references"][0] == (
"H3_CONTINUUM_AUDIO_REFERENCES" "H3_CONTINUUM_AUDIO_REFERENCES"
) )
mode_schema = v38["optional"]["video_reference_mode"]
assert mode_schema[0] == ("Follow Timeline", "Repeat Reference")
assert mode_schema[1]["default"] == "Repeat Reference"
assert v38.get("hidden") == v37.get("hidden") assert v38.get("hidden") == v37.get("hidden")
assert tuple(v38["required"])[-8:] == ( assert tuple(v38["required"])[-8:] == (
"generation_mode", "generation_mode",
+9 -2
View File
@@ -201,11 +201,18 @@ def test_v38_schema_keeps_legacy_aspect_and_appends_size_source_contract():
"the normal choice for T2VA or workflows without a First Image." "the normal choice for T2VA or workflows without a First Image."
), ),
} }
assert list(v38["optional"])[:-2] == list(v37["optional"]) assert list(v38["optional"])[:-3] == list(v37["optional"])
assert list(v38["optional"])[-2:] == ["audio_references", "image_references"] assert list(v38["optional"])[-3:] == [
"audio_references",
"image_references",
"video_reference_mode",
]
assert v38["optional"]["audio_references"][0] == ( assert v38["optional"]["audio_references"][0] == (
"H3_CONTINUUM_AUDIO_REFERENCES" "H3_CONTINUUM_AUDIO_REFERENCES"
) )
mode_schema = v38["optional"]["video_reference_mode"]
assert mode_schema[0] == ("Follow Timeline", "Repeat Reference")
assert mode_schema[1]["default"] == "Repeat Reference"
assert v38.get("hidden") == v37.get("hidden") assert v38.get("hidden") == v37.get("hidden")
@@ -0,0 +1,242 @@
"""CPU tests: real reference/grid helpers, deterministic recording VAE doubles.
These tests do not establish real GPU visual parity, memory peaks or frontend
compatibility. They can also run in the full repository's normal pytest suite.
"""
from __future__ import annotations
import copy
import dataclasses
import sys
import types
import pytest
import torch
from ComfyUI_H3_Continuum_Join import reference_video as reference
from ComfyUI_H3_Continuum_Join import temporal
from ComfyUI_H3_Continuum_Join import video_reference_modes as modes
class RecordingVAE:
def __init__(self):
self.inputs = []
def encode(self, frames):
self.inputs.append(frames.clone())
t = temporal.video_latent_t(int(frames.shape[0]))
# A CPU stand-in sufficient to verify the actual encoder's payload path.
return torch.full((1, 24, t, 2, 2), float(frames[0, 0, 0, 0]))
def frames(count=720, channels=3):
return torch.arange(count, dtype=torch.float32)[:, None, None, None].expand(count, 32, 32, channels).clone() / 1000
def source(count=720, seconds=5.0):
return modes.prepare_follow_video_source(
frames(count), chunk_seconds=seconds, output_width=32, output_height=32,
size_mode=reference.REFERENCE_VIDEO_SIZE_MATCH_OUTPUT,
)
def test_mode_is_optional_legacy_default_and_has_help():
options, definition = modes.mode_input_definition()
assert options == (modes.FOLLOW_TIMELINE, modes.REPEAT_REFERENCE)
assert definition["default"] == modes.REPEAT_REFERENCE
assert "Experimental" in definition["tooltip"]
@pytest.mark.parametrize("seconds", [5.0, 10.0, 15.0])
def test_follow_uses_natural_visible_frame_plan_not_queue_counter(seconds):
value = source(2000, seconds)
first = temporal.align_frame_count_up(round(seconds * 24))
retained = first
a = modes.select_follow_window(value, start_frame=0, visible_frames=first)
shape = temporal.make_extension_shape(22, (round(2 * seconds * 24) - retained) / 24)
b = modes.select_follow_window(value, start_frame=retained, visible_frames=shape.net_new_frames)
assert b.source_start == a.source_stop == retained
assert b.source_start != 0
assert b.available_frames == shape.total_frames - 22
# A future third group does not retroactively crop this source interval.
again = modes.select_follow_window(value, start_frame=retained, visible_frames=shape.net_new_frames)
assert again == b
@pytest.mark.parametrize("start,count", [(0, 243), (124, 238), (243, 119)])
def test_physical_terminal_window_is_not_split_or_context_shifted(start, count):
window = modes.select_follow_window(source(), start_frame=start, visible_frames=count)
assert window.source_start == start
assert window.source_stop == start + count
assert window.available_frames == count
@pytest.mark.parametrize("start,count", [(0, 124), (124, 119), (243, 238)])
def test_qwen_and_vae_receive_exactly_the_same_selected_interval(start, count):
value = source()
vae = RecordingVAE()
assets, window = modes.encode_follow_video_group(vae, value, start_frame=start, visible_frames=count)
sent = vae.inputs[0]
assert torch.equal(sent[:count], value.frames[start:start + count])
assert torch.equal(assets.item["data"], sent[::12])
assert assets.item["timestamps"] == [i / 2 for i in range(len(sent[::12]))]
assert assets.block["ref_audio_t"] == 0
assert assets.block["audio_latent"] is None
assert assets.block["latent"].device.type == "cpu"
assert sent.shape[0] == window.encoded_frames
@pytest.mark.parametrize("available", [5, 6, 21, 22, 23, 119, 120, 124, 237, 240, 243])
def test_only_short_native_grid_padding_never_looping(available):
value = source(available)
vae = RecordingVAE()
assets, window = modes.encode_follow_video_group(vae, value, start_frame=0, visible_frames=500)
sent = vae.inputs[0]
assert 0 <= sent.shape[0] - available <= 16
assert sent.shape[0] % 17 == 5
assert torch.equal(sent[:available], value.frames)
if sent.shape[0] > available:
assert torch.equal(sent[available:], value.frames[-1:].expand(sent.shape[0] - available, -1, -1, -1))
assert window.status == "partial_source"
@pytest.mark.parametrize("remaining", [0, 1, 2, 3, 4])
def test_exhausted_or_too_short_reference_continues_without_video(remaining):
value = source(120 + remaining)
vae = RecordingVAE()
calls = []
def builder(**kwargs):
calls.append(kwargs)
return {("same prompt", False): "text-only"}
cache, window = modes.build_follow_group_conditioning(
value, vae, start_frame=120, visible_frames=124,
conditioning_builder=builder, conditioning_kwargs={"prompts": ["same prompt"]},
)
assert calls[0]["timeline_video_assets"] is None
assert not vae.inputs
assert cache[("same prompt", False)] == "text-only"
assert window.status == ("source_exhausted" if remaining == 0 else "short_remainder_skipped")
def test_repeated_prompt_uses_separate_cache_and_payload_for_each_group():
value = source()
vae = RecordingVAE()
kwargs = {"prompts": ["the same text"]}
def builder(**request):
return {"the same text": request["timeline_video_assets"].item["data"]}
a, wa = modes.build_follow_group_conditioning(
value, vae, start_frame=0, visible_frames=124,
conditioning_builder=builder, conditioning_kwargs=kwargs,
)
b, wb = modes.build_follow_group_conditioning(
value, vae, start_frame=124, visible_frames=119,
conditioning_builder=builder, conditioning_kwargs=kwargs,
)
assert a is not b
assert wa.identity != wb.identity
assert a["the same text"][0, 0, 0, 0] != b["the same text"][0, 0, 0, 0]
assert kwargs == {"prompts": ["the same text"]} # no caller mutation
def test_reference_cache_key_contains_slice_and_respects_vae_identity(monkeypatch):
# Exercise the REAL encode_reference_video_cached function using a small
# cache-double, not a claim about the production cache's resource behavior.
class Cache:
def __init__(self):
self.items = {}
self.keys = []
def lookup(self, vae, key, event_sink=None):
self.keys.append((vae, key))
result = self.items.get((vae, key))
if event_sink: event_sink("hit" if result is not None else "miss")
return result
def supports_vae(self, vae):
return True
def store(self, vae, key, value, event_sink=None):
self.items[(vae, key)] = value
cache = Cache()
module = types.ModuleType("ComfyUI_H3_Continuum_Join.v3.ref_encode_cache")
module.get_ref_encode_cache = lambda: cache
module.make_ref_encode_cache_key = lambda *args: args
monkeypatch.setitem(sys.modules, module.__name__, module)
a, b = RecordingVAE(), RecordingVAE()
value = source()
for vae, start in ((a, 0), (a, 124), (a, 0), (b, 0)):
modes.encode_follow_video_group(vae, value, start_frame=start, visible_frames=124, cache_enabled=True)
assert len(a.inputs) == 2
assert len(b.inputs) == 1
assert cache.keys[0][1] == cache.keys[2][1]
assert cache.keys[0][1] != cache.keys[1][1]
assert cache.keys[0][0] is not cache.keys[3][0]
@pytest.mark.parametrize("n", [120, 124, 243])
def test_first_matching_interval_uses_legacy_encoder_with_exact_input_parity(n):
raw = frames(n)
old_source = reference.prepare_reference_video_source(raw, target_frames=temporal.align_frame_count_up(n), output_width=32, output_height=32, size_mode=reference.REFERENCE_VIDEO_SIZE_MATCH_OUTPUT)
old_vae, new_vae = RecordingVAE(), RecordingVAE()
old = reference.encode_reference_video(old_vae, old_source)
new, _ = modes.encode_follow_video_group(new_vae, source(n), start_frame=0, visible_frames=n)
assert torch.equal(old_vae.inputs[0], new_vae.inputs[0])
assert torch.equal(old.item["data"], new.item["data"])
assert old.item["timestamps"] == new.item["timestamps"]
assert torch.equal(old.block["latent"], new.block["latent"])
def test_contract_stable_for_extension_distinct_from_repeat():
value = source()
again = source()
assert value.contract == again.contract
assert "chunks" not in value.contract and "total_duration" not in value.contract
old = reference.prepare_reference_video_source(frames(), target_frames=124, output_width=32, output_height=32)
assert value.combined_hash != old.combined_hash
assert reference.combine_reference_video_identity("base", value) != reference.combine_reference_video_identity("base", old)
assert reference.combine_reference_video_identity("base", None) == "base"
def test_source_is_owned_cpu_snapshot_and_future_pixels_change_identity():
raw = frames()
value = modes.prepare_follow_video_source(raw, chunk_seconds=5, output_width=32, output_height=32)
saved = value.frames.clone()
raw[-1].fill_(0.9)
changed = modes.prepare_follow_video_source(raw, chunk_seconds=5, output_width=32, output_height=32)
assert torch.equal(value.frames, saved)
assert value.combined_hash != changed.combined_hash
assert value.frames.device.type == "cpu" and value.frames.dtype == torch.float32
def test_alpha_is_not_video_reference_content():
rgba = frames(22, 4)
rgba[..., 3] = 0.9
rgb = modes.prepare_follow_video_source(rgba[..., :3], chunk_seconds=5, output_width=32, output_height=32)
alpha = modes.prepare_follow_video_source(rgba, chunk_seconds=5, output_width=32, output_height=32)
assert alpha.combined_hash == rgb.combined_hash
assert torch.equal(alpha.frames, rgb.frames)
@pytest.mark.parametrize("start,count", [(-1, 20), (0, 0), (True, 2), (0, False), (0.2, 10), (0, 10.5)])
def test_invalid_internal_range_fails_without_sampling(start, count):
with pytest.raises(ValueError):
modes.select_follow_window(source(22), start_frame=start, visible_frames=count)
def test_invalid_frame_payload_detected():
with pytest.raises(reference.ReferenceVideoError):
modes.prepare_follow_video_source(torch.zeros(5, 2), chunk_seconds=5, output_width=32, output_height=32)
raw = frames(22)
raw[0, 0, 0, 0] = float("nan")
with pytest.raises(reference.ReferenceVideoError):
modes.prepare_follow_video_source(raw, chunk_seconds=5, output_width=32, output_height=32)
def test_metadata_is_plain_json_and_does_not_contain_tensors():
import json
value = source()
window = modes.select_follow_window(value, start_frame=124, visible_frames=119)
assert json.loads(json.dumps(window.contract)) == window.contract
assert json.loads(json.dumps(value.contract)) == value.contract
assert modes.FOLLOW_PLAN_KEY.startswith("_h3_continuum_")
def test_tests_do_not_initialize_cuda():
assert not torch.cuda.is_initialized()
+1
View File
@@ -16,6 +16,7 @@ EXPECTED_PUBLIC_NODE_DISPLAY_NAMES = {
"H3EasyLoadImage": "H3 Continuum Load Image", "H3EasyLoadImage": "H3 Continuum Load Image",
"H3EasyLoadAudio": "H3 Continuum Load Audio", "H3EasyLoadAudio": "H3 Continuum Load Audio",
"H3ContinuumLoadVideo": "H3 Continuum Load Video", "H3ContinuumLoadVideo": "H3 Continuum Load Video",
"H3ContinuumVideoAdapter": "H3 Continuum Video Adapter",
"H3ContinuumSecondPassV35": "H3 Continuum Second Pass", "H3ContinuumSecondPassV35": "H3 Continuum Second Pass",
"H3DecodeCacheHelper": "Decode Cache Helper", "H3DecodeCacheHelper": "Decode Cache Helper",
} }
+57 -5
View File
@@ -455,6 +455,10 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
RuntimeCoordinatorError, RuntimeCoordinatorError,
) )
from ..v3.sampling_engine import sample_physical_group from ..v3.sampling_engine import sample_physical_group
from ..video_reference_modes import (
FOLLOW_PLAN_KEY, FollowVideoSource, build_follow_group_conditioning,
)
following_video = isinstance(reference_video_source, FollowVideoSource)
storage_controller=get_active_run_storage() storage_controller=get_active_run_storage()
runtime_coordinator=InternalRuntimeCoordinator( runtime_coordinator=InternalRuntimeCoordinator(
storage_controller=storage_controller, storage_controller=storage_controller,
@@ -612,7 +616,11 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
if driving_audio_source is not None: if driving_audio_source is not None:
reuse_notes.insert(0,"Driving Audio: absolute-time guide slices enabled; original audio is preserved for final output.") reuse_notes.insert(0,"Driving Audio: absolute-time guide slices enabled; original audio is preserved for final output.")
if reference_video_source is not None: if reference_video_source is not None:
reuse_notes.insert(0,f"Video Reference conditioning: persistent across all chunks, 24 fps, resolved={reference_video_source.target_width}x{reference_video_source.target_height}, frames={reference_video_source.frame_count}.") if following_video:
reuse_notes.insert(0, "Timeline Video: Follow Timeline (Experimental), "
"24 fps; physical visible intervals; after source end, no video reference.")
else:
reuse_notes.insert(0,f"Video Reference conditioning: persistent across all chunks, 24 fps, resolved={reference_video_source.target_width}x{reference_video_source.target_height}, frames={reference_video_source.frame_count}.")
if reference_video_warning: reuse_notes.append(reference_video_warning) if reference_video_warning: reuse_notes.append(reference_video_warning)
if timeline_video_source is not None: if timeline_video_source is not None:
reuse_notes.insert(0,f"Timeline Video conditioning: chunk-local slices, size={timeline_video_source.size_mode}, resolved={timeline_video_source.target_width}x{timeline_video_source.target_height}.") reuse_notes.insert(0,f"Timeline Video conditioning: chunk-local slices, size={timeline_video_source.size_mode}, resolved={timeline_video_source.target_width}x{timeline_video_source.target_height}.")
@@ -677,6 +685,29 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
# metadata is changed in R3. # metadata is changed in R3.
storage_controller.execution_plan=execution_plan storage_controller.execution_plan=execution_plan
def follow_group_cache(group_prompt, include_last, start_frame, net_frames):
# A new prompt-cache dict per physical interval prevents equal-prompt
# chunks from accidentally reusing another interval's Qwen conditioning.
group_cache, window = build_follow_group_conditioning(
reference_video_source,
video_vae,
start_frame=int(start_frame),
visible_frames=int(net_frames),
conditioning_builder=_conditioning_cache,
conditioning_kwargs=dict(
clip=clip, prompts=[group_prompt], assets=assets,
final_has_last_frame=bool(include_last),
reference_assets=reference_assets,
reference_audio_assets=reference_audio_assets,
prompt_conditioning_cache=prompt_conditioning_cache,
prompt_cache_event=prompt_cache_events.append,
),
cache_enabled=reference_encode_cache,
cache_event=reference_encode_cache_events.append,
)
sampling_reports.append(window.report())
return group_cache, window.contract
cache={} cache={}
guide_assets=None guide_assets=None
if len(preserved)<chunks or capture_refine_context: if len(preserved)<chunks or capture_refine_context:
@@ -723,7 +754,7 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
if memory_collector is not None and driving_audio_source is not None: if memory_collector is not None and driving_audio_source is not None:
capture_memory(memory_collector, "finish_phase", phase="conditioning_driving_audio_vae") capture_memory(memory_collector, "finish_phase", phase="conditioning_driving_audio_vae")
reference_video_assets=None reference_video_assets=None
if reference_video_source is not None: if reference_video_source is not None and not following_video:
from ..reference_video import ( from ..reference_video import (
encode_reference_video, encode_reference_video,
encode_reference_video_cached, encode_reference_video_cached,
@@ -743,7 +774,7 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
) )
if memory_collector is not None: if memory_collector is not None:
capture_memory(memory_collector, "finish_phase", phase="conditioning_reference_video_vae") capture_memory(memory_collector, "finish_phase", phase="conditioning_reference_video_vae")
if timeline_video_source is None: if timeline_video_source is None and not following_video:
if memory_collector is not None: if memory_collector is not None:
capture_memory(memory_collector, "start_phase", phase="conditioning_prompt_clip") capture_memory(memory_collector, "start_phase", phase="conditioning_prompt_clip")
planned_normal_indices=tuple( planned_normal_indices=tuple(
@@ -829,6 +860,7 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
if memory_collector is not None: if memory_collector is not None:
capture_memory(memory_collector, "start_phase", phase="group_prepare", physical_group=sequence_index+1, logical_chunks=(sequence_index+1,)) capture_memory(memory_collector, "start_phase", phase="group_prepare", physical_group=sequence_index+1, logical_chunks=(sequence_index+1,))
prompt=prompts[sequence_index]; prompt_hash_value=prompt_hashes[sequence_index]; is_final=sequence_index==chunks-1; effective_reroll_nonce=int(reroll_nonce) if int(reroll_from_chunk)>0 and sequence_index+1>=int(reroll_from_chunk) else 0; seed=derive_chunk_seed(base_seed,sequence_index,effective_reroll_nonce); motion_score=0.0; video_context=None; audio_context=None; context_before=None; context_interop_emitted=False prompt=prompts[sequence_index]; prompt_hash_value=prompt_hashes[sequence_index]; is_final=sequence_index==chunks-1; effective_reroll_nonce=int(reroll_nonce) if int(reroll_from_chunk)>0 and sequence_index+1>=int(reroll_from_chunk) else 0; seed=derive_chunk_seed(base_seed,sequence_index,effective_reroll_nonce); motion_score=0.0; video_context=None; audio_context=None; context_before=None; context_interop_emitted=False
follow_slice_contract = None
timeline_video_assets=reference_video_assets timeline_video_assets=reference_video_assets
chunk_cache=cache chunk_cache=cache
if timeline_video_source is not None: if timeline_video_source is not None:
@@ -850,12 +882,21 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
if previous_state is None: if previous_state is None:
include_chunk_last=bool(last_frame is not None and is_final) include_chunk_last=bool(last_frame is not None and is_final)
conditioning_key=_conditioning_cache_key(prompt,include_last=include_chunk_last,reference_assets=reference_assets) conditioning_key=_conditioning_cache_key(prompt,include_last=include_chunk_last,reference_assets=reference_assets)
total_frames=initial_frame_count; latent=empty_h3_latent(width,height,total_frames); conditioning=attach_keyframes(chunk_cache[conditioning_key],frame_count=total_frames,first_latent=chunk_assets.first_latent,last_latent=chunk_assets.last_latent if include_chunk_last else None); clip_index=1; context_frames=0 total_frames=initial_frame_count; latent=empty_h3_latent(width,height,total_frames)
if following_video:
chunk_cache, follow_slice_contract = follow_group_cache(
prompt, include_chunk_last, retained_frames, total_frames,
)
conditioning=attach_keyframes(chunk_cache[conditioning_key],frame_count=total_frames,first_latent=chunk_assets.first_latent,last_latent=chunk_assets.last_latent if include_chunk_last else None); clip_index=1; context_frames=0
chunk_plan=make_plan(continuation=False,clip_index=clip_index,total_frames=total_frames,trim_frames=0,width=width,height=height,context_frames=5,state_capacity_frames=largest_context_capacity(total_frames),requested_extend_seconds=chunk_seconds,debug=debug); reason="initial clip" chunk_plan=make_plan(continuation=False,clip_index=clip_index,total_frames=total_frames,trim_frames=0,width=width,height=height,context_frames=5,state_capacity_frames=largest_context_capacity(total_frames),requested_extend_seconds=chunk_seconds,debug=debug); reason="initial clip"
else: else:
context_frames,motion_score,reason=choose_context_frames(continuity,previous_state); desired_cumulative=int(round((sequence_index+1)*chunk_seconds*FPS)); requested_new_frames=max(1,desired_cumulative-retained_frames); shape=make_extension_shape(context_frames,requested_new_frames/FPS); latent=empty_h3_latent(width,height,shape.total_frames) context_frames,motion_score,reason=choose_context_frames(continuity,previous_state); desired_cumulative=int(round((sequence_index+1)*chunk_seconds*FPS)); requested_new_frames=max(1,desired_cumulative-retained_frames); shape=make_extension_shape(context_frames,requested_new_frames/FPS); latent=empty_h3_latent(width,height,shape.total_frames)
include_chunk_last=bool(last_frame is not None and is_final) include_chunk_last=bool(last_frame is not None and is_final)
conditioning_key=_conditioning_cache_key(prompt,include_last=include_chunk_last,reference_assets=reference_assets) conditioning_key=_conditioning_cache_key(prompt,include_last=include_chunk_last,reference_assets=reference_assets)
if following_video:
chunk_cache, follow_slice_contract = follow_group_cache(
prompt, include_chunk_last, retained_frames, shape.net_new_frames,
)
continuation_first_latent=None if continuation_transport in (MASKED_VIDEO_PREFIX_V1,MASKED_AV_PREFIX_22_V1,MASKED_AV_PREFIX_39_V1) else chunk_assets.first_latent continuation_first_latent=None if continuation_transport in (MASKED_VIDEO_PREFIX_V1,MASKED_AV_PREFIX_22_V1,MASKED_AV_PREFIX_39_V1) else chunk_assets.first_latent
base_conditioning=attach_keyframes(chunk_cache[conditioning_key],frame_count=shape.total_frames,first_latent=continuation_first_latent,last_latent=chunk_assets.last_latent if include_chunk_last else None) base_conditioning=attach_keyframes(chunk_cache[conditioning_key],frame_count=shape.total_frames,first_latent=continuation_first_latent,last_latent=chunk_assets.last_latent if include_chunk_last else None)
if continuation_transport in (MASKED_VIDEO_PREFIX_V1,MASKED_AV_PREFIX_22_V1,MASKED_AV_PREFIX_39_V1): if continuation_transport in (MASKED_VIDEO_PREFIX_V1,MASKED_AV_PREFIX_22_V1,MASKED_AV_PREFIX_39_V1):
@@ -903,6 +944,8 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
conditioning=prepare_conditioning(base_conditioning,video_context=video_context,audio_context=audio_context,audio_grid_offset=grid_offset,context_frames=context_frames,new_frame_count=shape.total_frames,first_frame_policy=POLICY_REPLACE,preserve_last_frame=True) conditioning=prepare_conditioning(base_conditioning,video_context=video_context,audio_context=audio_context,audio_grid_offset=grid_offset,context_frames=context_frames,new_frame_count=shape.total_frames,first_frame_policy=POLICY_REPLACE,preserve_last_frame=True)
context_interop_emitted=True context_interop_emitted=True
clip_index=int(previous_state["clip_index"])+1; chunk_plan=make_plan(continuation=True,clip_index=clip_index,total_frames=shape.total_frames,trim_frames=context_frames,width=width,height=height,context_frames=context_frames,state_capacity_frames=largest_context_capacity(shape.net_new_frames),requested_extend_seconds=chunk_seconds,debug=debug) clip_index=int(previous_state["clip_index"])+1; chunk_plan=make_plan(continuation=True,clip_index=clip_index,total_frames=shape.total_frames,trim_frames=context_frames,width=width,height=height,context_frames=context_frames,state_capacity_frames=largest_context_capacity(shape.net_new_frames),requested_extend_seconds=chunk_seconds,debug=debug)
if follow_slice_contract is not None:
chunk_plan[FOLLOW_PLAN_KEY] = follow_slice_contract
if adaptive_observer_event is not None and not reusing_group: if adaptive_observer_event is not None and not reusing_group:
from ..v3.adaptive_continuity import INITIAL_TRANSPORT from ..v3.adaptive_continuity import INITIAL_TRANSPORT
adaptive_observer_event( adaptive_observer_event(
@@ -1106,8 +1149,15 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
) )
latent=empty_h3_latent(width,height,physical_frames) latent=empty_h3_latent(width,height,physical_frames)
conditioning_key=_conditioning_cache_key(prompt,include_last=True,reference_assets=reference_assets) conditioning_key=_conditioning_cache_key(prompt,include_last=True,reference_assets=reference_assets)
terminal_cache = cache
terminal_follow_slice = None
if following_video:
terminal_cache, terminal_follow_slice = follow_group_cache(
prompt, True, retained_frames,
physical_frames - physical_context_frames,
)
base_conditioning=attach_keyframes( base_conditioning=attach_keyframes(
cache[conditioning_key], terminal_cache[conditioning_key],
frame_count=physical_frames, frame_count=physical_frames,
first_latent=None if terminal_masked else assets.first_latent, first_latent=None if terminal_masked else assets.first_latent,
last_latent=assets.last_latent, last_latent=assets.last_latent,
@@ -1327,6 +1377,8 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
clip_index=physical_clip_index+offset clip_index=physical_clip_index+offset
context_frames=trim_frames if continuation else 5 context_frames=trim_frames if continuation else 5
chunk_plan=make_plan(continuation=continuation,clip_index=clip_index,total_frames=total_frames,trim_frames=trim_frames,width=width,height=height,context_frames=context_frames,state_capacity_frames=largest_context_capacity(total_frames-trim_frames),requested_extend_seconds=chunk_seconds,debug=debug) chunk_plan=make_plan(continuation=continuation,clip_index=clip_index,total_frames=total_frames,trim_frames=trim_frames,width=width,height=height,context_frames=context_frames,state_capacity_frames=largest_context_capacity(total_frames-trim_frames),requested_extend_seconds=chunk_seconds,debug=debug)
if terminal_follow_slice is not None:
chunk_plan[FOLLOW_PLAN_KEY] = dict(terminal_follow_slice)
logical_latent=latent_from_cpu(video_part,audio_part) logical_latent=latent_from_cpu(video_part,audio_part)
logical_seed=int(terminal_seed_plan["logical_entry_seeds"][offset]) logical_seed=int(terminal_seed_plan["logical_entry_seeds"][offset])
entry=make_chunk_entry(latent=logical_latent,plan=chunk_plan,prompt=prompts[sequence_index],prompt_hash=prompt_hashes[sequence_index],seed=logical_seed,context_frames=trim_frames,motion_score=motion_score,reused=False) entry=make_chunk_entry(latent=logical_latent,plan=chunk_plan,prompt=prompts[sequence_index],prompt_hash=prompt_hashes[sequence_index],seed=logical_seed,context_frames=trim_frames,motion_score=motion_score,reused=False)
+36 -9
View File
@@ -227,15 +227,26 @@ class H3ContinuumSamplerV34(H3ContinuumSamplerProduction):
target_frames=target_frames, target_frames=target_frames,
fps=FPS, fps=FPS,
) )
reference_video_source = prepare_reference_video_source( video_reference_mode = kwargs.pop("video_reference_mode", "Repeat Reference")
reference_video_1, if reference_video_1 is not None and video_reference_mode == "Follow Timeline":
target_frames=align_frame_count_up( from ..video_reference_modes import prepare_follow_video_source
int(round(float(kwargs["chunk_seconds"]) * FPS)) reference_video_source = prepare_follow_video_source(
), reference_video_1,
output_width=int(kwargs["width"]), chunk_seconds=float(kwargs["chunk_seconds"]),
output_height=int(kwargs["height"]), output_width=int(kwargs["width"]),
size_mode=str(video_reference_size), output_height=int(kwargs["height"]),
) size_mode=str(video_reference_size),
)
else:
reference_video_source = prepare_reference_video_source(
reference_video_1,
target_frames=align_frame_count_up(
int(round(float(kwargs["chunk_seconds"]) * FPS))
),
output_width=int(kwargs["width"]),
output_height=int(kwargs["height"]),
size_mode=str(video_reference_size),
)
outputs = super().run( outputs = super().run(
reference_audio_1=reference_audio_1, reference_audio_1=reference_audio_1,
reference_audio_vae=reference_audio_vae, reference_audio_vae=reference_audio_vae,
@@ -700,6 +711,22 @@ class H3ContinuumSamplerV38(H3ContinuumSamplerV37):
{"display_name": "Reference Images (Optional)", {"display_name": "Reference Images (Optional)",
"tooltip": "Optional Reference Images 4–9 from H3 Continuum Reference Images. Existing slots 1–3 remain first."}, "tooltip": "Optional Reference Images 4–9 from H3 Continuum Reference Images. Existing slots 1–3 remain first."},
) )
# Experimental, appended optional scalar: old API requests may omit it.
# Existing required widget positions and all IMAGE sockets are unchanged.
from ..video_reference_modes import mode_input_definition
optional["video_reference_mode"] = mode_input_definition()
video_definition = optional["reference_video_1"]
video_options = dict(video_definition[1])
video_options.update(
display_name="Timeline Video Frames",
tooltip=(
"24 fps IMAGE frames from the Video Adapter. Follow Timeline "
"selects successive physical output intervals; Repeat Reference "
"keeps the legacy shared-prefix guide. Source audio is not included. "
"Follow is Experimental; this is not a hard motion-copy control."
),
)
optional["reference_video_1"] = (video_definition[0], video_options)
schema["optional"] = optional schema["optional"] = optional
return schema return schema
+299
View File
@@ -0,0 +1,299 @@
"""Experimental Follow Timeline over the existing 24-fps IMAGE input.
Repeat Reference deliberately continues through reference_video.py unchanged.
A Follow source has a different identity, but its identity does not depend on
Chunks. A physical window is selected from the existing retained-frame plan,
not a Queue-local counter. No sampling, continuation or audio code lives here.
"""
from __future__ import annotations
from dataclasses import dataclass
import math
from typing import Any, Callable
import torch
from .reference_video import (
REFERENCE_VIDEO_PREPROCESS_VERSION,
REFERENCE_VIDEO_SIZE_EFFICIENT,
REFERENCE_VIDEO_SIZE_OPTIONS,
ReferenceVideoAssets,
ReferenceVideoError,
ReferenceVideoSource,
_canonical_hash,
_prepare_reference_frames,
encode_reference_video,
encode_reference_video_cached,
resolve_reference_video_size,
)
from .temporal import align_frame_count_up
REPEAT_REFERENCE = "Repeat Reference"
FOLLOW_TIMELINE = "Follow Timeline"
VIDEO_REFERENCE_MODES = (FOLLOW_TIMELINE, REPEAT_REFERENCE)
FOLLOW_VERSION = 1
FPS = 24
FOLLOW_PLAN_KEY = "_h3_continuum_video_reference_v1"
WINDOW_POLICY = "natural_visible_frames_excluding_context_v1"
END_POLICY = "remaining_interval_then_no_reference_v1"
def mode_input_definition() -> tuple:
"""Append this widget; never insert it into the old serialized prefix."""
return (
VIDEO_REFERENCE_MODES,
{
"default": REPEAT_REFERENCE,
"display_name": "Video Reference Mode",
"tooltip": (
"Experimental. Follow Timeline selects the source interval for each "
"physical generation group, then continues without video reference "
"after the source ends. Repeat Reference preserves the previous "
"Video Guide behavior: the same bounded prefix for every group. "
"This is independent of Prompt Format. Input frames must be 24 fps."
),
},
)
@dataclass(frozen=True)
class FollowVideoSource:
"""Run-scoped, normalized source. Do not mutate frames after preparation."""
frames: torch.Tensor
source_shape: tuple[int, ...]
source_dtype: str
source_sha256: str
size_mode: str
target_width: int
target_height: int
chunk_seconds: float
combined_hash: str
@property
def frame_count(self) -> int:
return int(self.source_shape[0])
@property
def contract(self) -> dict[str, Any]:
# No configured Chunks or total duration: adding future groups must not
# invalidate an otherwise unchanged saved prefix. Mode is nevertheless
# part of the reference identity, so Follow and Repeat never collide.
return {
"follow_video_contract_version": FOLLOW_VERSION,
"mode": FOLLOW_TIMELINE,
"source_shape": list(self.source_shape),
"source_dtype": self.source_dtype,
"source_sha256": self.source_sha256,
"source_fps": FPS,
"size_mode": self.size_mode,
"target_width": self.target_width,
"target_height": self.target_height,
"chunk_seconds": self.chunk_seconds,
"preprocess_version": REFERENCE_VIDEO_PREPROCESS_VERSION,
"window_policy": WINDOW_POLICY,
"end_policy": END_POLICY,
"combined_hash": self.combined_hash,
}
@dataclass(frozen=True)
class FollowVideoWindow:
"""Half-open source frame range, before H3's short alignment padding."""
requested_start: int
requested_stop: int
source_start: int
source_stop: int
encoded_frames: int
source_frames: int
identity: str
@property
def available_frames(self) -> int:
return self.source_stop - self.source_start
@property
def active(self) -> bool:
return self.available_frames >= 5
@property
def status(self) -> str:
if self.available_frames == 0:
return "source_exhausted"
if not self.active:
return "short_remainder_skipped"
if self.source_stop < self.requested_stop:
return "partial_source"
return "active"
@property
def contract(self) -> dict[str, Any]:
return {
"version": FOLLOW_VERSION,
"mode": FOLLOW_TIMELINE,
"requested_start_frame": self.requested_start,
"requested_stop_frame": self.requested_stop,
"source_start_frame": self.source_start,
"source_stop_frame": self.source_stop,
"source_frames": self.source_frames,
"encoded_frames": self.encoded_frames,
"padding_frames": self.encoded_frames - self.available_frames if self.active else 0,
"fps": FPS,
"status": self.status,
"slice_sha256": self.identity,
}
def report(self) -> str:
if not self.active:
return (
f"Timeline Video: output start={self.requested_start / FPS:.3f}s; "
f"{self.status}; no video reference for this physical group."
)
return (
"Timeline Video: source "
f"[{self.source_start / FPS:.3f}, {self.source_stop / FPS:.3f})s "
f"({self.available_frames} frames, H3={self.encoded_frames}); {self.status}."
)
def prepare_follow_video_source(
frames: torch.Tensor,
*,
chunk_seconds: float,
output_width: int,
output_height: int,
size_mode: str = REFERENCE_VIDEO_SIZE_EFFICIENT,
) -> FollowVideoSource:
"""Keep the full normalized source for future windows, without a GPU copy.
This IMAGE-input implementation is not a lazy media-file reader. Its CPU
source copy is intentional: hashing and encoding must see the same pixels.
It makes no claim of reducing long-input RAM use.
"""
if not torch.is_tensor(frames) or frames.ndim != 4 or int(frames.shape[-1]) < 3:
raise ReferenceVideoError("Timeline Video Frames requires IMAGE [T,H,W,C] with RGB")
if int(frames.shape[1]) < 1 or int(frames.shape[2]) < 1:
raise ReferenceVideoError("Timeline Video Frames has empty spatial dimensions")
seconds = float(chunk_seconds)
if not math.isfinite(seconds) or seconds <= 0:
raise ValueError("Invalid internal chunk duration for Timeline Video")
mode = str(size_mode) if str(size_mode) in REFERENCE_VIDEO_SIZE_OPTIONS else REFERENCE_VIDEO_SIZE_EFFICIENT
normalized, shape, dtype, fingerprint = _prepare_reference_frames(
frames, frame_count=int(frames.shape[0])
)
width, height = resolve_reference_video_size(
int(shape[2]), int(shape[1]),
output_width=int(output_width), output_height=int(output_height), size_mode=mode,
)
descriptor = {
"follow_video_contract_version": FOLLOW_VERSION,
"mode": FOLLOW_TIMELINE,
"source_shape": list(shape),
"source_dtype": dtype,
"source_sha256": fingerprint,
"source_fps": FPS,
"size_mode": mode,
"target_width": width,
"target_height": height,
"chunk_seconds": seconds,
"preprocess_version": REFERENCE_VIDEO_PREPROCESS_VERSION,
"window_policy": WINDOW_POLICY,
"end_policy": END_POLICY,
}
return FollowVideoSource(
normalized, shape, dtype, fingerprint, mode, width, height, seconds,
_canonical_hash(descriptor),
)
def select_follow_window(
source: FollowVideoSource, *, start_frame: int, visible_frames: int,
) -> FollowVideoWindow:
"""Select the existing physical plan's visible range, never its context.
Use the natural (pre-final-crop) output range. Clamping this to the current
requested total would change an old final group's conditioning on extension.
Only the available source and H3's <=16-frame grid padding limit the window.
"""
if isinstance(start_frame, bool) or isinstance(visible_frames, bool):
raise ValueError("Physical video window indices must be integers")
if int(start_frame) != start_frame or int(visible_frames) != visible_frames:
raise ValueError("Physical video window indices must be integers")
start, count = int(start_frame), int(visible_frames)
if start < 0 or count < 1:
raise ValueError("Invalid internal visible-frame window")
stop = start + count
first = min(start, source.frame_count)
last = min(stop, source.frame_count)
available = last - first
encoded = align_frame_count_up(available) if available >= 5 else 0
identity = _canonical_hash({
"source": source.combined_hash,
"requested_start": start,
"requested_stop": stop,
"source_start": first,
"source_stop": last,
"encoded_frames": encoded,
"window_policy": WINDOW_POLICY,
})
return FollowVideoWindow(start, stop, first, last, encoded, source.frame_count, identity)
def encode_follow_video_group(
video_vae: Any,
source: FollowVideoSource,
*,
start_frame: int,
visible_frames: int,
cache_enabled: bool = False,
cache_event: Callable | None = None,
) -> tuple[ReferenceVideoAssets | None, FollowVideoWindow]:
window = select_follow_window(source, start_frame=start_frame, visible_frames=visible_frames)
if not window.active:
return None, window
selected = source.frames[window.source_start:window.source_stop]
# Both the Qwen item and native Video VAE latent are made by the SAME
# existing encoder from this one selected frame batch. Slice identity is
# included in the old cache's source.combined_hash slot.
selected_source = ReferenceVideoSource(
frames=selected,
source_shape=tuple(int(v) for v in selected.shape),
source_dtype=source.source_dtype,
source_sha256=window.identity,
size_mode=source.size_mode,
target_width=source.target_width,
target_height=source.target_height,
frame_count=window.encoded_frames,
combined_hash=window.identity,
)
if cache_enabled:
assets = encode_reference_video_cached(video_vae, selected_source, cache_event=cache_event)
else:
assets = encode_reference_video(video_vae, selected_source)
return assets, window
def build_follow_group_conditioning(
source: FollowVideoSource,
video_vae: Any,
*,
start_frame: int,
visible_frames: int,
conditioning_builder: Callable,
conditioning_kwargs: dict[str, Any],
cache_enabled: bool = False,
cache_event: Callable | None = None,
) -> tuple[dict, FollowVideoWindow]:
"""Use a fresh prompt-cache dictionary per physical reference interval.
Never reuse a prior group's text conditioning merely because prompt text
matches. The VAE cache remains reusable because its key includes the slice.
"""
assets, window = encode_follow_video_group(
video_vae, source, start_frame=start_frame, visible_frames=visible_frames,
cache_enabled=cache_enabled, cache_event=cache_event,
)
kwargs = dict(conditioning_kwargs)
kwargs["timeline_video_assets"] = assets
return conditioning_builder(**kwargs), window
+15
View File
@@ -1340,6 +1340,20 @@ function configureIntuitiveV38Ux(node) {
setWidgetVisible(findWidget(node, LEGACY_RUN_NAME_WIDGET), advanced && storageEnabled); setWidgetVisible(findWidget(node, LEGACY_RUN_NAME_WIDGET), advanced && storageEnabled);
setWidgetVisible(findWidget(node, REROLL_NONCE_WIDGET), advanced && explicitRegeneration); setWidgetVisible(findWidget(node, REROLL_NONCE_WIDGET), advanced && explicitRegeneration);
applyV38WidgetHelp(node); applyV38WidgetHelp(node);
// This is the appended native widget, not another transient proxy.
// Core persists its value; visibility never changes the saved mode.
const videoMode = findWidget(node, "video_reference_mode");
setWidgetVisible(videoMode, videoConnected);
if (videoMode) {
videoMode.label = "Video Reference Mode";
setWidgetTooltip(videoMode, String(videoMode.value) === "Follow Timeline"
? "Experimental: use successive 24 fps source intervals. After the source ends, continue without a video reference. Independent of Prompt Format."
: "Legacy Video Guide behavior: reuse the same bounded source prefix on each physical group. This is not source looping or stretching.");
}
const timelineInput = node.inputs?.find((item) => item.name === "reference_video_1");
if (timelineInput) {
timelineInput.label = "Timeline Video Frames";
}
moveFacadeWidgetsToFront(node, orderedNames); moveFacadeWidgetsToFront(node, orderedNames);
node.__h3ContinuumProductionUxRefresh?.(); node.__h3ContinuumProductionUxRefresh?.();
node.setDirtyCanvas?.(true, true); node.setDirtyCanvas?.(true, true);
@@ -1348,6 +1362,7 @@ function configureIntuitiveV38Ux(node) {
for (const [name, key] of ( for (const [name, key] of (
[ [
["prompt_mode", "__h3ContinuumFacadePromptMode"], ["prompt_mode", "__h3ContinuumFacadePromptMode"],
["video_reference_mode", "__h3ContinuumFacadeVideoReferenceMode"],
["continuity", "__h3ContinuumFacadeContinuity"], ["continuity", "__h3ContinuumFacadeContinuity"],
["base_seed", "__h3ContinuumFacadeBaseSeed"], ["base_seed", "__h3ContinuumFacadeBaseSeed"],
["audio_continuity", "__h3ContinuumFacadeAudioContinuity"], ["audio_continuity", "__h3ContinuumFacadeAudioContinuity"],
+4 -1
View File
@@ -17,7 +17,10 @@ export function normalizeReferenceAudioLabels(node) {
return true; return true;
} }
for (const input of node.inputs || []) { for (const input of node.inputs || []) {
const label = NORMALIZED_INPUT_LABELS[input.name]; const label = input.name === "reference_video_1"
&& node.comfyClass === "H3ContinuumSamplerV38"
? "Timeline Video Frames"
: NORMALIZED_INPUT_LABELS[input.name];
if (label) { if (label) {
input.label = label; input.label = label;
} }
+42
View File
@@ -0,0 +1,42 @@
import { app } from "../../scripts/app.js";
const NODE_CLASS = "H3ContinuumSamplerV38";
const MODE = "video_reference_mode";
const REPEAT = "Repeat Reference";
const FOLLOW = "Follow Timeline";
export function restoreVideoReferenceMode(node, info) {
const widget = node.widgets?.find((item) => item.name === MODE);
if (!widget) return;
const linked = info?.inputs?.some((input) => (
(input.name === MODE || input.widget?.name === MODE) && input.link != null
));
if (linked) return; // The connected input, not a hidden widget, owns its value.
let value;
const named = info?.widgets_values_named;
if (named && Object.prototype.hasOwnProperty.call(named, MODE)) {
value = named[MODE];
} else if (!named) {
// The existing V38X2 backend widgets occupy slots 0..29. This new
// optional scalar is appended; no upload/UI/transient row is inserted.
const values = info?.widgets_values;
if (Array.isArray(values) && values.length > 30) value = values[30];
}
// In particular A(Follow)->B(old workflow) must not inherit A's mode.
widget.value = value === FOLLOW ? FOLLOW : REPEAT;
node.__h3ContinuumIntuitiveUxRefresh?.();
}
app.registerExtension({
name: "H3Continuum.TimelineVideoModeExperimental",
beforeRegisterNodeDef(nodeType, nodeData) {
if (nodeData?.name !== NODE_CLASS && nodeType?.prototype?.comfyClass !== NODE_CLASS) return;
const previous = nodeType.prototype.onConfigure;
nodeType.prototype.onConfigure = function(info, ...args) {
const result = previous?.call(this, info, ...args);
restoreVideoReferenceMode(this, info);
return result;
};
},
});