Minimax_H3_Face_Resample

This is the middle node of the pack's face-fix trio and the only one that actually generates anything. Minimax_H3_Face_Cut hands it uniform res × res crop rows; this node runs a block-level img2img pass over those crops with the same MiniMax H3 model, VAE and text encoder your main sampler used; Minimax_H3_Face_Blend pastes the result back.

Mechanically it's an ADetailer-style detected-crop re-render - the loop the detailing doc describes as zoom in on a mask, upscale, diffuse, composite. What's specific here is that the "crop" is a video with chunk seams in it, so the crop rows are split into blocks that mirror your main sampler's segmentation, each block gets inter-block continuation anchoring, and each block's encoding is padded to the 17n+5 frame grid and trimmed after decode. The refinement inherits the same temporal structure as the thing it's repairing instead of fighting it.

The two wiring modes are the part people trip over. In integrated mode you connect info from the main H3 sampler; the node then pulls model/VAE/CLIP out of that info and ignores its own model/vae/clip ports entirely. In standalone mode you leave info empty and wire the local model/VAE/CLIP yourself - which is how you use this on ordinary video that H3 never touched, and it's a genuinely useful escape hatch for repairing faces in anything. The parameter port is required in both modes. It's not the main sampler's info; don't cross the wires, and specifically don't connect Face_Cut's shot_info here - that node explicitly says it doesn't consume it.

Inputs that matter

Two required inputs are just plumbing: crop_images and face_pack, both straight off Face_Cut. Beyond those and the two wiring-mode inputs (info, or model/vae/clip):

• sigmas - required, both modes. A decreasing σ schedule ending at 0, normally from a scheduler node. This is your denoise: the starting σ is how much noise gets added to the crop, so a lower start is a gentler fix. SplitSigmas' low output works (it ends at 0), and raising its step softens the repair - just keep the step under your total step count, or the low side shrinks to a bare 0 and refinement degenerates into a pass-through. Newer ComfyUI builds have SplitSigmasDenoise, which splits by denoise value directly and is less fiddly.
• face_prompt - a face-specific prompt, with a real job. Your scene prompts were written for whole-shot generation and often say nothing about faces; at high σ they actively mislead the redraw. Use this to state what you actually want.
• prompt_mode - how that combines with the original per-segment prompt. prepend keeps the scene context and your reference declarations. replace_text keeps only the reference tags. replace_all drops the segment prompt entirely. For high-σ structural repair, replace_text is the author's first choice.
• identity_ref (default true) - the main anti-jitter lever. It picks the sharpest crop row per identity and injects it as a <Picture N> reference for that identity's redraw, so the face doesn't reinvent itself between blocks. Making things worse? Just turn it off.
• seed - keep control_after_generate on fixed. Randomise it and the row cache will never hit.
• color_match (default true) - Reinhard-style per-subtrack colour match back to the source crops so the repaired face sits in its surroundings. Applied after the cache loads, so toggling it is free.
• sampler_name / scheduler and the optional sampler port are standalone-mode only - integrated mode gets the sampler from info, and plugging a sampler in overrides both dropdowns.
• Optional autogrow ref_images / ref_videos / ref_video_audios - only passed when declared in the segment prompt with a Picture tag. Index subtlety: the auto identity anchor is <Picture 1> with no refs connected, and shifts to "connected refs + 1" when you have some - two ref images in means your auto anchor is <Picture 3>.
• Optional audio matters only in images mode, where the pack has no audio ledger: connect the source audio and it's encoded with audio_vae for lipsync. With neither, you get a silent placeholder and a warning.

Outputs: images (the refined canvas, row-for-row matching crop_images) → Face_Blend's canvas. bbox (the repackaged geometry pack) → Face_Blend's bbox. identity_refs previews the anchors actually used this run - check it, because a wrong anchor means the wrong identity gets repaired and you'll only find out three nodes later.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/supElement/ComfyUI_MinimaxH3_AutoContext.git
cd ComfyUI_MinimaxH3_AutoContext
pip install -r requirements.txt

Or search ComfyUI_MinimaxH3_AutoContext in Manager, then restart. Either way you need the H3 model, video VAE, audio VAE and CLIP - integrated mode pulls them out of info, standalone mode makes you wire them yourself. "Standalone" means the video wasn't made by this sampler; the repair still runs on H3.

Common issues

"Resample reports missing inputs." Two causes, one per mode: integrated mode without info connected, or standalone mode without model + vae + clip. The parameter port being connected doesn't count - it's required in both modes but is not the wiring-mode switch.

The face stays stubbornly identical to the original. Your σ start is too low, usually from a too-aggressive SplitSigmas step. Same if you went looking for a denoise slider: there isn't one, σ is it.

An obvious hard seam at a chunk boundary. Try prompt_mode: replace_text for high σ, and keep color_match on. If it persists, the repair is too aggressive for that window - lower the σ start.

Identity drifts across blocks. This is what identity_ref exists for. Confirm with its identity_refs output that the anchor is the right person and not a blurry frame.

Nothing changes when you re-run with different settings. Cache. Changing the σ schedule invalidates automatically; changing prompts, seeds or reference declarations does too. When in doubt, tick clear_cache or delete ComfyUI/output/cache/node_<node id>/. The pack caps that directory at 32 GB and evicts oldest-first (H3_CACHE_MAX_GB tunes it, 0 disables).