Create H3 RefMod

This is the node people actually mean when they say the pack is "LoRAs without training." MiniMax H3 is genuinely good at reference conditioning - the community reaction to its motion control was "like Kling" - but every native reference video injects thousands of tokens into the packed sequence, which is why people started asking whether you could compress that. Extract H3 RefMod turns one or more stills or clips of a concept into a RefMod: a small .safetensors containing a latent that still rides H3's native ref path, at a fraction of the compute.

And here's the part that surprises people: there is no diffusion model involved and no training. Extraction only needs the H3 video VAE. The heaviest thing in this workflow is the encode.

How it works

Plug stills into the autogrow ref_image_1, video frames into ref_video_1 (a + button adds more slots of each type). Each ref is a view of the same concept. Refs get stacked along the latent time axis into one video-kind mod, so a multi-image moodboard keeps each image's own content instead of averaging it into mush.

Then pick a mode:

• training (default) - resize to ref_resolution, VAE-encode, then average-pool to a small grid (pool_h × pool_w, auto-fit to the source's aspect ratio so a portrait isn't squished square) and refine the small latent against the full encode with gradient steps. 16×16 = 64 tokens per frame, which is the pack's "concept sweet spot"; 32×32 = 256; 64×64 = 1024 (full-mode parity).
• encode - store the straight full-res encode at ref_resolution. Max identity, but an MB-sized mod at ~1K tokens per 1024px image.

The identity dial is refinement steps, not a lesson in who the person is - the docstring is explicit that it doesn't teach the model which attribute to keep. 500 is the default, 0 is pure pooling. Small grids lose faces and fine detail; that's the whole trade.

Inputs a beginner actually sets

Honestly: name, mode, ref_resolution, the pool dials, identity, and max_tokens. max_tokens (default 5120) is your safety valve - if stacked refs overflow it, near-duplicate latent frames get dropped first, then frames are resampled. If you know you'll only ever use one concept per file, concept_type and description are optional metadata that also feed the loaders' prompt_hint output. Connect a subject mask with background_retention when encode mode keeps pulling in background you don't own. merge (training only) refines one consensus latent against every ref - good for a folder of angles; motion_only is an experiment for capturing motion without appearance.

Wire the mods output straight into Apply H3 RefMod for an instant test, or set save and pick the file back up later with Load H3 RefMods (the name shows in its dropdown after a reload). Save goes to the first registered refmods root - models/refmods by default, or wherever you've mapped it in extra_model_paths.yaml - with an optional subfolder.

Install

ComfyUI Manager (search "ComfyUI-MiniMaxH3Mod") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Luisacaotica/ComfyUI-MiniMaxH3Mod

Restart. Deps are light (safetensors, numpy, Pillow). You need the H3 weights and its video VAE already working in a current ComfyUI - the pack explicitly doesn't import ComfyUI-MiniMaxH3 - and H3's licence is territory-restricted (US, EU, UK, Korea excluded).

Common issues

Don't expect magic from a small grid: the README calls the "8×8 concept / 16×16 identity" sweet spot folklore, not a validated split. If a face refuses to come through, go bigger on the grid or switch to encode - pooling is lossy in exactly the way that destroys facial identity. Fast motion is the pack's known hard case: a burst of hand signs smears into slow softness, so extract only the moments that matter and describe the motion in your prompt. And remember extraction and voice are separate: appearance lives here, audio lives in Extract H3 Audio RefMod or the Master node.