MiniMax H3 Reference Video Fade (Experimental)

H3's reference video is strong - sometimes too strong. It doesn't reduce a Ref2VA clip to poses; it VAE-encodes the complete 24 fps reference and packs its latent rows into the same transformer sequence, so a character-swap prompt can reproduce the reference's camera, color, background, and texture along with the motion you actually wanted. That's why a simple swap "clones" the source so convincingly, and why you might wish you could loosen its grip. You can't just drop every other frame - that changes speed and timing. This experimental MODEL patch instead keeps the complete video and fades its attention value contribution late in denoising.

What it does

MiniMaxH3ReferenceVideoFadeModelPatch patches the H3 model so that during early denoising, native reference-video rows stay at full strength, then after fade_start their value contribution follows a half-cosine curve down to end_strength. It targets every native video and video_audio reference block, scaling H3's V slices in place - no sequence-sized mask, no duplicate model, no second attention pass, so the memory cost is minimal.

The presets tell the story:

• full - unchanged behavior.
• balanced - full through 67% of denoising, fading to 20%.
• freer - full through 50%, fading to 15%.
• early_only - full through 50%, fading to zero.
• custom - your own custom_fade_start / custom_end_strength.

progress is measured against the complete sigma schedule, which is why the optional full_sigmas input exists: in split-sigma sampling, stage two needs the original full schedule so it continues the same fade instead of starting a fresh local curve. Use the same schedule on every switched model branch.

Still pictures, Qwen presentation tokens, reference audio, continuation guides, and the target AV rows are untouched - this is a smooth reduction of the video ref's influence, not a hard removal, and its Q/K rows still carry weaker semantic influence.

Install and gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

Restart, or install via ComfyUI Manager under "MiniMax H3 Contex Loop". No pip dependencies; a current ComfyUI with native Add Guide (PR #15439) is expected. Models aren't bundled - and yes, check the MiniMax H3 Community License territory restriction.

It's experimental for a reason: tune presets against a test render, and don't expect early_only to behave like a hard "remove the reference after 50%" - it's still a smooth curve, and identity that lives only in the reference video can still leak in early denoising. It composes with an existing SolAttn or Comfy Kitchen attention override, so if you already run those, place this patch after your backend patches and route its output through every sampler stage.