Node packVerificadoExterno

ComfyUI-MAINodes

MatlowAI's MiniMax-H3 ComfyUI nodes: Contact-Sheet diffusion + Motion Lab (test-time de-roping of fast motion)

Fuentegithub.com/matlowai/ComfyUI-MAINodes

Responsable@matlowaiA · adoptioncalidad

Perfil del paquete

Estado: Verificado
Publicador: @matlowai
Fuente: GitHub

Léeme

ComfyUI-MAINodes

Custom nodes for MiniMax-H3. Two groups: Motion Lab (a test-time fix for fast-motion smearing) and Contact-Sheet diffusion (five views of one subject from one reference image).

Like it? A star helps. Want to feed the GPU? Sponsor the experiments.

Start here

Which graph: examples/README.md. Four starting points, the essential node chain as a diagram, and what each specialised graph adds to it.
Which settings for your card: HARDWARE.md. VRAM and system RAM rows, measured.
Which dial for which symptom: TUNING.md.
What is still alpha: ALPHA.md.

The short version: start from examples/motion_pipeline_ref2va_audioinit.json, keep pass 1 at 12 steps on the base model, let the turbo LoRA do pass 2, and reach for inject (0.50 sharper, 0.70 safer) before any other dial.

What's new

Cloned before 2026-08-19 and git pull errors? Run git fetch origin && git reset --hard origin/main, or re-clone.

Repair, a per-block clock, and the chaining nodes (2026-08-24, alpha). Five additions, all opt-in and all alpha. H3 Repair Plan / H3 Repair Splice re-render a few bad frames of a finished clip and hand back the same clip with only those frames different (below). H3 DyRoPE gives the true-duration RoPE clock to SOME blocks or SOME steps instead of all of them, which is what separates the timing correction from the flash and jitter it used to cost (below). H3 V2V Init grew audio_prefix_ticks, the audio twin of the video prefix freeze, so a continuation can carry the previous segment's sound across the join (below). H3 Drift Control and H3 Delta Color Carry attack the two things that decay down a chain of segments, clean-conditioning error and VAE round-trip colour (below). And window mode, the excerpt-shaped graph where only part of a long clip is retimed, finally has a page: docs/WINDOW_MODE.md.

Long de-ropes on small cards (2026-08-18, alpha). A de-rope pass at d_max 4 on an 8 to 12 second clip is ~200k packed tokens, and the stock H3 block materialises its fused QKV and SwiGLU tensors for the whole sequence (8.6 and 15.4 GiB at that length), which is what OOMs 24 GB cards and, at 1376x768, a 96 GB one. H3 Streamed Blocks runs every DiT block in token chunks with the same math (int8 and W4A8 activations quantise per row and accumulate in int32, so the result is bit-equal to stock; measured on same-seed renders, video and audio) and streams the output head the same way. With ComfyUI's dynamic VRAM and --fast-disk, the same 702-frame pass that OOMed a 96 GB card renders on a 16 GB card in a 32 GB machine at the same seconds per step (316 vs 311), because the step is attention-bound at that length and the weight traffic hides under it. Alongside it: H3 Memory Probe (per-block, per-phase VRAM and RSS ledger with a hoverable timeline, plus an optional allocator trace), H3 Free Cache (return the allocator pool before VAE decode; 17 GiB back on the long pass), H3 Evict Text Encoder. All opt-in, nothing changes unless the node is in the graph. The memory numbers behind them are in the docstrings of vram_lab.py; the small-card recipe, what to expect and the environment it was measured in are in LOWVRAM.md, with the example graph examples/motion_pipeline_lowvram.json; the alpha status is in ALPHA.md. Evening additions: two opt-in K/V store options in the same node (kvi8r, rotated int8 K/V, forward peak +9.5 GiB instead of +11.9; kvi8s, the same bytes in SageAttention's layout attended on int8/fp8 tensor cores with no dequant), both approximation tier with the measurements and what is still owed in LOWV