chris.dumas/comfyui-workflows
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Refresh MiniMax H3 Seamless Chain v2.6.0

Browse Source
This commit is contained in:
Morpheus
2026-08-17 18:07:58 +00:00
parent dd81085967
commit bbdbd2129d
2 changed files with 3496 additions and 3754 deletions
Download Patch File Download Diff File Expand all files Collapse all files
models/minimax/minimax-h3-seamless-chain-v2/README.md
+9 -3
View File
@@ -1,10 +1,11 @@
# MiniMax H3 Seamless Chain v2
# MiniMax H3 Seamless Chain v2.6.0
## Summary
Same-day `MiniMax-H3` multishot workflow from `joeygambino`. This is the full
graph, not the CORE cut-down version: it adds the LLM prompt lane, remote text
encoder support, raw-latent continuity, and the speed/VRAM helper switches.
encoder support, raw-latent continuity, the extend-take path, and the
speed/VRAM helper switches.
## Model Family
@@ -19,7 +20,7 @@ encoder support, raw-latent continuity, and the speed/VRAM helper switches.
- Workflow pack page: <https://huggingface.co/joeygambino/MiniMax-H3-Multishot-Workflow>
- Workflow file: <https://huggingface.co/joeygambino/MiniMax-H3-Multishot-Workflow/resolve/main/workflows/H3_Seamless_Chain_v2.json?download=true>
- Original publisher: `joeygambino`
- Date imported: `2026-08-16`
- Date imported: `2026-08-17`
## Developer notes
@@ -27,6 +28,9 @@ encoder support, raw-latent continuity, and the speed/VRAM helper switches.
block cache, custom scripts, and GGUF support on top of the base H3 pack.
- The workflow is built around the `ComfyUI-H3-Multishot` node pack and the
same MiniMax-H3 weight family used by the CORE workflow.
- The current upstream release is v2.6.0 and adds the long-chain audio drift
fix, `audio_tone_control=flatten`, `take_seconds`, `audio_pin_frames`, and
the `H3_Extend_Take` workflow.
- The prompt-writer lane expects a `qwen3:14b`-class local or remote
OpenAI-compatible endpoint if you keep it enabled.
@@ -36,6 +40,8 @@ encoder support, raw-latent continuity, and the speed/VRAM helper switches.
feature set and are willing to pay for it in complexity.
- On an `RTX 3070 8 GB`, this is a heavy workflow, but the memory and remote
encoder options are the kind of practical extras that make it worth keeping.
- The v2.6.0 audio fixes make this the strongest MiniMax H3 roundup item
today.
- If you want the simpler starting point, the CORE workflow is still the
cleaner local baseline.
models/minimax/minimax-h3-seamless-chain-v2/workflow.json
+104 -368
View File
@@ -2,7 +2,7 @@
"id": "h3-seamless-chain-v1",
"revision": 0,
"last_node_id": 69,
"last_link_id": 95,
"last_link_id": 98,
"nodes": [
{
"id": 28,
@@ -726,57 +726,6 @@
"image"
]
},
{
"id": 14,
"type": "H3AnySwitch",
"pos": [
-1362.3228245783803,
1056.1629296808258
],
"size": [
159.7818359375,
66
],
"flags": {
"collapsed": true
},
"order": 55,
"mode": 0,
"inputs": [
{
"name": "use_on",
"type": "BOOLEAN",
"link": 14
},
{
"name": "off_path",
"shape": 7,
"type": "*",
"link": 12
},
{
"name": "on_path",
"shape": 7,
"type": "*",
"link": 13
}
],
"outputs": [
{
"name": "value",
"type": "*",
"slot_index": 0,
"links": [
90
]
}
],
"title": "gate: block_cache",
"properties": {
"Node name for S&R": "H3AnySwitch"
},
"widgets_values": []
},
{
"id": 10,
"type": "H3AnySwitch",
@@ -870,8 +819,7 @@
"type": "*",
"slot_index": 0,
"links": [
11,
12
90
]
}
],
@@ -921,6 +869,11 @@
"name": "num_frames"
},
"link": 26
},
{
"name": "refs_attached",
"type": "BOOLEAN",
"link": 97
}
],
"outputs": [
@@ -975,12 +928,8 @@
"id": 51,
"type": "ImageBatch",
"pos": [
-2582.009842829537,
1136.0292081407952
],
"size": [
140,
46
-2582,
1110
],
"flags": {
"collapsed": true
@@ -1020,8 +969,8 @@
"id": 49,
"type": "LoadImage",
"pos": [
-2583.4862919122725,
1261.2350651000709
-2583,
1290
],
"size": [
300.0083984375,
@@ -1060,8 +1009,8 @@
"id": 50,
"type": "LoadImage",
"pos": [
-2583.5884958767797,
1610.1483776240293
-2584,
1640
],
"size": [
302.8521484375,
@@ -1096,62 +1045,12 @@
"image"
]
},
{
"id": 57,
"type": "ImageBatch",
"pos": [
-2204.315668416005,
1132.1128904695718
],
"size": [
234.61484375,
46
],
"flags": {
"collapsed": true
},
"order": 63,
"mode": 0,
"inputs": [
{
"name": "image1",
"type": "IMAGE",
"link": 77
},
{
"name": "image2",
"type": "IMAGE",
"link": 78
}
],
"outputs": [
{
"name": "IMAGE",
"type": "IMAGE",
"slot_index": null,
"links": [
79
]
}
],
"title": "auto refs 1+2",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.32.0",
"Node name for S&R": "ImageBatch"
},
"widgets_values": []
},
{
"id": 60,
"type": "PrimitiveBoolean",
"pos": [
-2239.719668751282,
1201.2431022131907
],
"size": [
472.335546875,
58
-2240,
1136
],
"flags": {},
"order": 7,
@@ -1163,11 +1062,12 @@
"type": "BOOLEAN",
"slot_index": null,
"links": [
83
83,
97
]
}
],
"title": "USE AUTO REFS (off = the two manual LoadImage slots)",
"title": "USE AUTO REFS",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.32.0",
@@ -1177,62 +1077,12 @@
false
]
},
{
"id": 58,
"type": "ImageBatch",
"pos": [
-2202.246562478545,
1166.043343639111
],
"size": [
234.61484375,
46
],
"flags": {
"collapsed": true
},
"order": 64,
"mode": 0,
"inputs": [
{
"name": "image1",
"type": "IMAGE",
"link": 79
},
{
"name": "image2",
"type": "IMAGE",
"link": 80
}
],
"outputs": [
{
"name": "IMAGE",
"type": "IMAGE",
"slot_index": null,
"links": [
81
]
}
],
"title": "auto refs +3",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.32.0",
"Node name for S&R": "ImageBatch"
},
"widgets_values": []
},
{
"id": 56,
"type": "H3AutoRefs",
"pos": [
-2247.312640975518,
1299.0918739796418
],
"size": [
481.7271484375,
354
-2247,
1329
],
"flags": {},
"order": 62,
@@ -1242,6 +1092,11 @@
"name": "prompt_text",
"type": "STRING",
"link": 76
},
{
"name": "extra_text",
"type": "STRING",
"link": 98
}
],
"outputs": [
@@ -1249,25 +1104,19 @@
"name": "ref_1",
"type": "IMAGE",
"slot_index": null,
"links": [
77
]
"links": null
},
{
"name": "ref_2",
"type": "IMAGE",
"slot_index": null,
"links": [
78
]
"links": null
},
{
"name": "ref_3",
"type": "IMAGE",
"slot_index": null,
"links": [
80
]
"links": null
},
{
"name": "ref_4",
@@ -1316,9 +1165,16 @@
"type": "STRING",
"slot_index": null,
"links": null
},
{
"name": "refs_batch",
"type": "IMAGE",
"links": [
96
]
}
],
"title": "AUTO REFS - scans the script for character folders under refs_root",
"title": "AUTO REFS (one folder per character under refs_root)",
"properties": {
"Node name for S&R": "H3AutoRefs"
},
@@ -1552,8 +1408,8 @@
"id": 53,
"type": "Note",
"pos": [
-2246.1739069358314,
1697.188077432785
-2246,
1735
],
"size": [
481.6705703188454,
@@ -1569,7 +1425,7 @@
"text": ""
},
"widgets_values": [
"CHARACTER PHOTOS (optional)\n\nDrop in up to 9 photos of your character and they'll look like that in EVERY\nshot. More angles = better. Works best with a ref2va model file.\n\nTwo different people? Set reference_subjects: \"3,3\" means photos 1-3 are\nperson A, photos 4-6 are person B. Skip this and faces will blend."
"CHARACTER PHOTOS (optional)\n\nTWO WAYS IN - pick one:\n\nAUTO: switch USE AUTO REFS on. That alone is enough - it also tells the writer (refs_attached) to stop describing the face and point at the photographs instead; measured: a written identity sentence overrides the photographs. Put one folder per character under input/h3_refs/ (folder name = the name used in the script) and the photos are found and sent to the sampler.\n\nMANUAL: leave USE AUTO REFS off, load photos into the two REFERENCE slots and switch the MANUAL REFS gate on. If the writer is on, also feed its refs_attached input a true boolean (or the writer will describe a face that fights your photos).\n\nThey'll look like that in EVERY shot. More angles = better. Works best with a ref2va model file.\n\nTwo different people? Set reference_subjects on the sampler: \"3,3\" means photos 1-3 are person A, photos 4-6 are person B. Skip this and faces will blend."
],
"color": "#432",
"bgcolor": "#653"
@@ -1606,8 +1462,8 @@
-545.0694242301886
],
"size": [
420.9913637192858,
286.69295482646885
421,
330
],
"flags": {},
"order": 12,
@@ -1617,7 +1473,7 @@
"title": "START HERE",
"properties": {},
"widgets_values": [
"START HERE\n\nThis workflow turns a list of shot descriptions into ONE video with sound.\n\n1. Type your shots in the REWRITER box (or use the prompt files switch).\n2. Check MASTER CONTROLS: size, seconds per shot, number of shots.\n3. Press Run. One video comes out, already joined.\n\nDefaults set for 16-24 GB cards. Got more VRAM? Raise size to 736x1280.\n\nEvery panel below is OPTIONAL, the workflow runs fine with all of them off."
"START HERE\n\nThis workflow turns a list of shot descriptions into ONE video with sound.\n\n1. Type your shots in the REWRITER box (or use the prompt files switch).\n2. Check MASTER CONTROLS: size, seconds per shot, number of shots.\n3. Press Run. One video comes out, already joined.\n\nDefaults set for 16-24 GB cards at 736x1280. Under 16 GB? Drop to 640x1152 (keep people at or above ~1 MP).\n\nWant one person talking for a minute straight? Open H3_Extend_Take (the main workflow) or set take_seconds on MASTER CONTROLS here.\n\nGuides: the 5-minute guide + Every setting explained - links in the read-me note to the LEFT.\n\nEvery panel below is OPTIONAL, the workflow runs fine with all of them off."
],
"color": "#432",
"bgcolor": "#653"
@@ -1768,7 +1624,7 @@
"title": "VRAM / SPEED - check these switches",
"properties": {},
"widgets_values": [
"VRAM / SPEED - check these switches\n\nHover any switch on this panel or on the SPEED BOOSTERS node for what it does and what it costs.\n\nThis panel ships all OFF - that's the tested setup. The one worth knowing: remote_encoder moves the text encoder's 15+ GB to a second PC (see the REMOTE ENCODER note).\n\nFaster renders? The SPEED BOOSTERS node next to the sampler: block cache ships ON (~11% faster, no visible change). Spectrum (~29%), TeaCache (~14-21%) and EasyCache (~21%) are faster still but can distort people - test on your own scene first, and never stack Spectrum with TeaCache.\n\nRenders slow or crashing? Lower the size or seconds-per-shot in MASTER CONTROLS first. That fixes 9 out of 10 memory problems."
"VRAM / SPEED - check these switches\n\nHover any switch for what it does and what it costs.\n\nThis panel ships all OFF - that's the tested setup. sol_attn and chunk_ffn lower peak VRAM (install ComfyUI-sol-attn, un-bypass the two patch nodes with Ctrl+B, then flip). remote_encoder moves the text encoder's 15+ GB to a second PC (see the REMOTE ENCODER note).\n\nFaster renders? The SPEED BOOSTERS node next to the sampler. All ship OFF: Spectrum (~29%), TeaCache (~14-21%) and EasyCache (~21%) are real speed but can distort people - test on your own scene first, and never stack Spectrum with TeaCache. Block cache does nothing at 14 steps (measured 0 hits) - only worth it at 30+ steps.\n\nRenders slow or crashing? Lower the size or seconds-per-shot in MASTER CONTROLS first. That fixes 9 out of 10 memory problems."
],
"color": "#432",
"bgcolor": "#653"
@@ -1818,53 +1674,6 @@
8192
]
},
{
"id": 13,
"type": "MiniMaxH3BlockCacheT8",
"pos": [
-1732.2582272281413,
753.4415810341072
],
"size": [
545.0306686636656,
202
],
"flags": {},
"order": 54,
"mode": 4,
"inputs": [
{
"name": "model",
"type": "MODEL",
"link": 11
}
],
"outputs": [
{
"name": "MODEL",
"type": "MODEL",
"slot_index": 0,
"links": [
13
]
}
],
"title": "needs comfyui-minimax-h3-blockcache-T8",
"properties": {
"aux_id": "T8mars/comfyui-minimax-h3-blockcache-T8",
"ver": "3cd968f11a0b717683b6abb0477c43fe175cb608",
"Node name for S&R": "MiniMaxH3BlockCacheT8"
},
"widgets_values": [
0.12,
0.08,
0.95,
2,
"cpu",
8,
false
]
},
{
"id": 9,
"type": "MiniMaxH3MemoryEfficientSolAttentionPatch",
@@ -1955,11 +1764,7 @@
"type": "H3StudioSwitches",
"pos": [
-2579.4819954135974,
216.53003324327636
],
"size": [
387.7721060280252,
323.57544854310277
262
],
"flags": {},
"order": 19,
@@ -1967,15 +1772,13 @@
"inputs": [],
"outputs": [
{
"name": "two_pass_upscale",
"name": "two_pass (removed)",
"type": "BOOLEAN",
"slot_index": 0,
"links": []
"links": null
},
{
"name": "sol_attn",
"type": "BOOLEAN",
"slot_index": 1,
"links": [
6
]
@@ -1983,44 +1786,36 @@
{
"name": "chunk_ffn",
"type": "BOOLEAN",
"slot_index": 2,
"links": [
10
]
},
{
"name": "spectrum",
"name": "spectrum (see Speed Boosters)",
"type": "BOOLEAN",
"slot_index": 3,
"links": []
"links": null
},
{
"name": "block_cache",
"name": "block_cache (see Speed Boosters)",
"type": "BOOLEAN",
"slot_index": 4,
"links": [
14
]
"links": null
},
{
"name": "dual_clock_sampler",
"name": "dual_clock (removed)",
"type": "BOOLEAN",
"slot_index": 5,
"links": []
"links": null
},
{
"name": "hybrid_cond",
"name": "hybrid_cond (removed)",
"type": "BOOLEAN",
"slot_index": 6,
"links": []
"links": null
},
{
"name": "remote_encoder",
"type": "BOOLEAN",
"links": [
95
],
"slot_index": 7
]
}
],
"title": "VRAM / SPEED SWITCHES (all OFF = verified)",
@@ -2028,11 +1823,6 @@
"Node name for S&R": "H3StudioSwitches"
},
"widgets_values": [
false,
false,
false,
false,
false,
false,
false,
false
@@ -2175,8 +1965,9 @@
false,
false,
0.15,
true,
false
false,
false,
0.12
]
},
{
@@ -2235,10 +2026,10 @@
"mode": 0,
"inputs": [],
"outputs": [],
"title": "joeygambino Workflow Pack v.2.2.6",
"title": "MiniMax-H3 Multishot v2.6.0 - links",
"properties": {},
"widgets_values": [
"## MiniMax-H3 Seamless Chain v2.5.2\n\n**Updates & guides**\n- [Civitai page + articles](https://civitai.com/models/2833322)\n- [HuggingFace workflow repo](https://huggingface.co/joeygambino/MiniMax-H3-Multishot-Workflow)\n- [GitHub](https://github.com/jlucasmcrell/ComfyUI-H3-Multishot)\n\n**Model files (DiT)**\n- [curve GGUFs](https://huggingface.co/joeygambino/MiniMax-H3-curve-GGUF)\n- [comfy-native ref2va](https://civitai.com/models/2857809)\n\n**Text encoder + VAEs**\n- [encoder GGUF](https://huggingface.co/joeygambino/MiniMax-H3-encoder-GGUF)\n- [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3)\n\nNodes: search **H3 Multishot** in ComfyUI-Manager\n"
"## MiniMax-H3 Seamless Chain v2.6.0\n\n**Guides**\n- [The 5-minute guide](https://civitai.com/articles/34047/make-talking-videos-with-minimax-h3-the-5-minute-guide-26)\n- [Every setting explained (deep manual)](https://civitai.com/articles/34046/every-setting-explained-the-seamless-chain-deep-manual)\n\n**Updates**\n- [Civitai page](https://civitai.com/models/2833322)\n- [HuggingFace workflow repo](https://huggingface.co/joeygambino/MiniMax-H3-Multishot-Workflow)\n- [GitHub](https://github.com/jlucasmcrell/ComfyUI-H3-Multishot)\n\n**Model files (DiT)**\n- [curve GGUFs](https://huggingface.co/joeygambino/MiniMax-H3-curve-GGUF)\n- [comfy-native ref2va](https://civitai.com/models/2857809)\n\n**Text encoder + VAEs**\n- [encoder GGUF](https://huggingface.co/joeygambino/MiniMax-H3-encoder-GGUF)\n- [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3)\n\nNodes: search **H3 Multishot** in ComfyUI-Manager\n"
],
"color": "#322",
"bgcolor": "#533",
@@ -2298,7 +2089,7 @@
"title": "BEFORE YOU UN-BYPASS",
"properties": {},
"widgets_values": [
"EXTRA SPEED SWITCHES (advanced)\n\nThese two nodes need extra packs installed first - links in SETTINGS.md.\nWithout the packs, leave them alone (bypassed) and everything works."
"EXTRA VRAM SWITCHES (advanced)\n\nThese two nodes lower peak VRAM (memory-efficient attention, chunked feed-forward) and need ComfyUI-sol-attn installed first - link in SETTINGS.md. Install, un-bypass them (Ctrl+B), then turn sol_attn / chunk_ffn ON in the VRAM / SPEED SWITCHES panel. Without the pack, leave them alone (bypassed) and everything works."
],
"color": "#432",
"bgcolor": "#653"
@@ -2431,12 +2222,8 @@
"id": 52,
"type": "H3OptionalImage",
"pos": [
-2582.534018230245,
1168.2892898420184
],
"size": [
324.255153067057,
58
-2582,
1186
],
"flags": {},
"order": 61,
@@ -2446,7 +2233,7 @@
"name": "image",
"shape": 7,
"type": "IMAGE",
"link": 84
"link": 82
}
],
"outputs": [
@@ -2454,11 +2241,11 @@
"name": "image",
"type": "IMAGE",
"links": [
75
84
]
}
],
"title": "REFERENCE gate (OFF = no references)",
"title": "MANUAL REFS gate",
"properties": {
"Node name for S&R": "H3OptionalImage",
"cnr_id": "comfy-core",
@@ -2472,12 +2259,8 @@
"id": 59,
"type": "H3AnySwitch",
"pos": [
-1997.6519652191153,
1151.6287036397023
],
"size": [
188.51328125,
66
-1940,
1215
],
"flags": {
"collapsed": true
@@ -2494,13 +2277,13 @@
"name": "off_path",
"shape": 7,
"type": "*",
"link": 82
"link": 84
},
{
"name": "on_path",
"shape": 7,
"type": "*",
"link": 81
"link": 96
}
],
"outputs": [
@@ -2509,7 +2292,7 @@
"type": "*",
"slot_index": null,
"links": [
84
75
]
}
],
@@ -2558,7 +2341,8 @@
"type": "*",
"slot_index": 0,
"links": [
67
67,
98
]
}
],
@@ -3113,38 +2897,6 @@
0,
"BOOLEAN"
],
[
11,
12,
0,
13,
0,
"*"
],
[
12,
12,
0,
14,
1,
"*"
],
[
13,
13,
0,
14,
2,
"MODEL"
],
[
14,
3,
4,
14,
0,
"BOOLEAN"
],
[
19,
20,
@@ -3435,7 +3187,7 @@
],
[
75,
52,
59,
0,
30,
7,
@@ -3449,52 +3201,12 @@
0,
"STRING"
],
[
77,
56,
0,
57,
0,
"IMAGE"
],
[
78,
56,
1,
57,
1,
"IMAGE"
],
[
79,
57,
0,
58,
0,
"IMAGE"
],
[
80,
56,
2,
58,
1,
"IMAGE"
],
[
81,
58,
0,
59,
2,
"IMAGE"
],
[
82,
51,
0,
59,
1,
52,
0,
"IMAGE"
],
[
@@ -3507,10 +3219,10 @@
],
[
84,
59,
0,
52,
0,
59,
1,
"IMAGE"
],
[
@@ -3555,7 +3267,7 @@
],
[
90,
14,
12,
0,
66,
0,
@@ -3600,6 +3312,30 @@
68,
0,
"BOOLEAN"
],
[
96,
56,
11,
59,
2,
"IMAGE"
],
[
97,
60,
0,
6,
3,
"BOOLEAN"
],
[
98,
46,
0,
56,
1,
"STRING"
]
],
"groups": [
@@ -3718,7 +3454,7 @@
-2602.009400625792,
1069.657936214502,
856.613073901633,
874.4434561814621
914
],
"color": "#3f789e",
"flags": {}