Files
comfyui-workflows/optimized/R2V/minimax/continuum-tiled/README.md
T
2026-10-19 11:16:56 +00:00

11 KiB
Raw Blame History

MiniMax H3 Continuum — tiled upscale / multi-clip prompt study

This folder contains the supplied MiniMax H3 Continuum workflow with the DUMAS tiled latent upscale second pass. It is the working candidate for testing prompt direction and multi-clip construction toward a 20-second final video.

Source and status

  • Source: Unsaved_Workflow_5_1.json, supplied 2026-10-14.
  • Status: imported baseline; smoke validation pending on the live server.
  • Workflow: workflow.json
  • References: character-reference.png and scene-reference.png
  • Custom nodes: DUMAS H3 Continuum, DUMAS H3 Continuum Tiled, and the native MiniMax H3/Fearnworks/PlagueKind dependencies used by the graph.

What is fixed

These values are held constant while testing prompts and clip scaling:

  • Model: fastvideo_fasth3_8step_v2_pruned_int8_convrot.safetensors
  • Text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
  • Video VAE: minimax_h3_video_vae_int8_convrot.safetensors
  • Audio VAE: minimax_h3_audio_vae_fp32.safetensors
  • Main sampler: euler_ancestral
  • Main scheduler: simple
  • Main steps: 8
  • Main denoise: 1.0
  • Main seed: 12
  • Tiled second-pass sampler: sa_solver
  • Tiled second-pass scheduler: simple
  • Tiled second-pass steps: 3 for the adjusted EasyCache + clear-VRAM flows (the original control workflows remain at 2)
  • Tiled second-pass denoise: 0.2
  • Tiled second-pass CFG: 1.0
  • Final output: 1920×1088
  • Latent upscale model: minimax_h3_latent_upscaler_3d_conv_v1_fp16.safetensors
  • Tile geometry: 1056×640, 96-pixel overlap, 32-pixel fade
  • Tiled mode: enabled through DUMAS H3 Continuum Second Pass
  • Generation mode: Full Run so all configured chunks are generated in one queue action; Review Each Chunk is reserved for manual continuation tests.

The imported graph starts at two 4-second Continuum chunks. The test matrix in tests.json increases the number of chunks through 20, 24, and 28 seconds, then tests a six-chunk, five-second-per-chunk configuration for the 30-second target. No sampler, scheduler, CFG, upscale or tile changes are included in this study.

Prompt strategy

The same character and scene references, seed, camera language, and visual direction are retained across the matrix. Prompt tests change only the structure and specificity of the action plan:

  1. Compact chronological action beats.
  2. Detailed chronological beats with explicit continuity.
  3. Detailed beats with camera direction and physical state tracking.
  4. Full 20-second plan with five linked 4-second beats.
  5. Longform 30-second plan with six linked 5-second beats.

Test procedure

  1. Run the 2-chunk smoke row first.
  2. Confirm the final output is routed through the DUMAS Second Pass with tiled_upscale=true and reaches 1920×1088.
  3. Repeat with the next prompt variant using the same 2 chunks.
  4. Keep the best prompt fixed, then test 3, 4, and 5 chunks.
  5. Continue through 24s and 28s; only then attempt the 30s five-second-chunk configuration.
  6. Stop if the server errors, OOMs, or runtime growth becomes disproportionate.
  7. Record each result beside the matrix with prompt hash, chunk count, runtime, output dimensions, and a motion/continuity verdict.

Test matrix and results

workflow.json is the supplied graph. workflow-spectrum-bypassed.json is an explicit structural bypass: the sampler, scheduler, second pass, and decode path connect directly to the LoRA/model output rather than relying only on Spectrum's enabled=false switch.

ID Workflow Prompt Chunks × seconds Target Status Result
CT-01 Spectrum graph Compact 2 × 4 8s configured Completed 17m 49.959s render; expected 4.0s output, 1920×1088, 96 frames
CT-02 Spectrum + EasyCache + clear VRAM Detailed continuity 2 × 4 8s configured Completed 21m 49.127s render; expected 4.0s output, 1920×1088, 96 frames
CT-03 Spectrum + EasyCache + clear VRAM Cinematic camera 2 × 4 8s configured Completed 22m 50.266s render; expected 4.0s output, 1920×1088, 96 frames
CT-04 Spectrum + EasyCache + clear VRAM Best smoke prompt 3 × 4 12s Completed 71m 39.693s render; 12.0s output, 1920×1088, 288 frames; 3/3 chunks
CT-05 Spectrum + EasyCache + clear VRAM Best smoke prompt 4 × 4 16s Lost Prompt absent from queue and history; no output available
CT-06 Spectrum + EasyCache + clear VRAM Best smoke prompt 5 × 4 20s Completed 121m 47.263s; 20.0s output, 1920×1088, 480 frames; 5/5 chunks
CT-07 Spectrum + EasyCache + clear VRAM Best smoke prompt 6 × 5.17 (124 frames) 31.02s Planned —
CT-08 Spectrum + EasyCache + clear VRAM Best smoke prompt 7 × 7.29 (175 frames) 51.03s Planned —
CT-09 Spectrum + EasyCache + clear VRAM Longform 30s 6 × 10.13 (243 frames) 60.78s Planned —
CT-B01 Spectrum bypassed Compact 2 × 4 8s configured Completed 16m 42.576s render; 4.0s output, 1920×1088, 24 fps, 96 frames; 1/2 logical chunks emitted
CT-B02 Bypassed + EasyCache + clear VRAM Detailed continuity 2 × 3.75 (90 frames) 7.5s Planned —
CT-B03 Bypassed + EasyCache + clear VRAM Cinematic camera 2 × 3.75 (90 frames) 7.5s Planned —
CT-B04 Bypassed + EasyCache + clear VRAM Best smoke prompt 3 × 5.17 (124 frames) 15.51s Planned —
CT-B05 Bypassed + EasyCache + clear VRAM Best smoke prompt 4 × 5.17 (124 frames) 20.68s Planned —
CT-B06 Bypassed + EasyCache + clear VRAM Best smoke prompt 5 × 5.17 (124 frames) 25.85s Planned —
CT-B07 Bypassed + EasyCache + clear VRAM Best smoke prompt 6 × 5.17 (124 frames) 31.02s Planned —
CT-B08 Bypassed + EasyCache + clear VRAM Best smoke prompt 7 × 5.17 (124 frames) 36.19s Planned —
CT-B09 Bypassed + EasyCache + clear VRAM Longform 30s 6 × 7.29 (175 frames) 43.74s Planned —
CT-10 Spectrum + EasyCache + clear VRAM Best smoke prompt 2 × 3.75 (90 frames) 7.5s Planned —
CT-11 Spectrum + EasyCache + clear VRAM Best smoke prompt 2 × 5.17 (124 frames) 10.34s Planned —
CT-12 Spectrum + EasyCache + clear VRAM Best smoke prompt 2 × 7.29 (175 frames) 14.58s Planned —
CT-13 Spectrum + EasyCache + clear VRAM Best smoke prompt 2 × 10.13 (243 frames) 20.26s Completed 80m 30.304s; 20.25s output, 1920×1088, 486 frames; 2/2 chunks
CT-14 Spectrum + EasyCache + clear VRAM Four 243-frame chunks 4 × 10.13 (243 frames) 40.52s Lost Prompt absent from ComfyUI queue and history; no output available; no duplicate submission made
CT-B10 Bypassed + EasyCache + clear VRAM Best smoke prompt 2 × 3.75 (90 frames) 7.5s Planned —
CT-B11 Bypassed + EasyCache + clear VRAM Best smoke prompt 2 × 5.17 (124 frames) 10.34s Planned —
CT-B12 Bypassed + EasyCache + clear VRAM Best smoke prompt 2 × 7.29 (175 frames) 14.58s Planned —
CT-B13 Bypassed + EasyCache + clear VRAM Best smoke prompt 2 × 10.13 (243 frames) 20.26s Planned —
CT-B14 Bypassed + EasyCache + clear VRAM Longform 30s 2 × 15.08 (362 frames) 30.16s Planned —
CT-15 Spectrum + EasyCache + clear VRAM 226-frame chunks 2 × 9.4 18.8s Lost Prompt absent from history and queue; no output available

The CT-10–CT-14 sweep isolates individual chunk length: it holds the chunk count at two and uses the valid maximum-frame choices 90, 124, 175, 243, and 362 (3.75s, 5.17s, 7.29s, 10.13s, and 15.08s). The bypass counterparts use the same settings. This shows whether longer individual clips are viable before combining them into longer projects.

Runtime planning

The latest observed reference is 18 minutes for one 4-second chunk. Until more measurements are available, the monitoring estimate is:

expected minutes = 18 × (total generated seconds ÷ 4) + 5-minute buffer

Examples:

Generated duration Estimated render Monitoring window
4s 18m 23m
5s 22.5m 28m
8s 36m 41m
20s 90m 95m
30s 135m 140m

These are planning windows, not performance guarantees. Each completed test must replace its estimate with the measured runtime and the model should be recalibrated if multi-clip scaling is not approximately linear.

CT-01 and CT-B01 remain the two no-cache controls. Every later row adds EasyCache with reuse threshold 0.3, active from 20% to 90% of sampling, and verbose logging disabled. Those rows also pass the completed DUMAS second-pass video output through AusBoss Free Memory 🆎 immediately before decode. It clears ComfyUI/CUDA caches while preserving the video-sample wire. The clear-VRAM node is deliberately after tiled upscale, so it cannot evict the model before the second pass has finished.

The bypass series uses the identical seed, references, prompt text, model, samplers, scheduler, upscale settings, tile geometry, and chunk plan. Only the Spectrum path changes. Results should be compared in pairs: CT-01 vs CT-B01, CT-02 vs CT-B02, and so on.

CT-15 frame mapping

The H3 temporal grid requires frame counts of 17k + 5. 226 is valid, and the node’s 0.1-second precision maps it to chunk_seconds=9.4: round(9.4 × 24) = 226. CT-15 therefore tests two 226-frame chunks with an expected assembled duration of approximately 18.8 seconds.

CT-B01 measured relationship

CT-B01 was accepted with dumas-continuum-character.png and dumas-continuum-scene.png, and configured for 2 logical chunks × 4.0s = 8.0s. The review-mode execution completed only the first physical chunk, so node 382 emitted a 4.0s MP4 (96 frames at 24 fps) while logical chunk 2 remained pending. The measured render duration was 1,002.576s (16m 42.576s), and the output metadata reported 1920×1088 at 24 fps.

Files

  • workflow.json — supplied ComfyUI API graph.
  • workflow-spectrum-bypassed.json — explicit Spectrum-bypass comparison graph.
  • workflow-easy-cache.json — Spectrum graph with EasyCache enabled.
  • workflow-spectrum-bypassed-easy-cache.json — bypass graph with EasyCache enabled.
  • workflow-easy-cache-clear-vram.json — Spectrum graph with EasyCache and pass-through VRAM cleanup after the DUMAS second pass.
  • workflow-spectrum-bypassed-easy-cache-clear-vram.json — bypass graph with EasyCache and pass-through VRAM cleanup after the DUMAS second pass.
  • prompt.txt — controlled base prompt for the study.
  • prompt-variants.json — exact smoke-test and 30-second prompt variants.
  • variables.json — immutable constants and allowed test variables.
  • tests.json — ordered test matrix.
  • character-reference.png, scene-reference.png — fixed references.
  • results/ — reserved for labelled outputs and metadata. Server output uses the matching continuum-tiled/<test-id>/ prefix.