Nougan RefMod Apply

MiniMax H3's workflow has a natural seam: MiniMax H3 Image To Video produces positive conditioning, and everything downstream - injecting the references, weighting them, sequencing them - is bookkeeping on top of it. The upstream custom-node route made you connect a loader to an applier to get a character into a clip. Nougan RefMod Apply collapses that: pick the character on the node, conditioning in, conditioning out.

This is the node the author tells you to use, and it's the right default. Reach for Nougan RefMod Loader only when you need mods as data for something custom.

What you wire

conditioning (required) - the positive conditioning from H3 Image To Video. The tooltip is specific about that, and it matters: this isn't a general-purpose conditioning filter.

Then three slots, mod_1 (required, (none) allowed), mod_2, mod_3, each with the same five controls:

• visual_strength_N - face/likeness. 0 skips loading the visuals entirely.
• audio_strength_N - voice. 0 skips the audio.
• strength_N - master multiplier applied on top of both.
• copies_N (1–10) - repeats the reference. Each copy costs its full tokens in every block, so this is a token bill, not a free emphasis control.
• components_N - All, Visual or Audio. Unused parts are never read from disk.

That separation is H3's modal design showing through. Face and voice are independent, so a slot can contribute a likeness with no voice, or a voice with no face, or both at half strength. If no character is selected the node passes conditioning through untouched and says so in the console, which makes it safe to leave in a graph.

Outputs

conditioning goes to your sampler. prompt_hint is a merged string of the selected refs' concept types and descriptions - concatenate it onto your positive prompt through a string node rather than retyping the character's description. info is the ledger: each ref, its strength, its token count, the total. Keep a Preview Any on it while you tune; it answers "why is this render so slow" faster than anything else.

Why this exists at all

Baking a character into a refmod isn't training - Nougan RefMod Bake 🧬 is VAE extraction plus a container, no gradients. What it buys you is reuse: portraits encoded once, a name in a dropdown, consistent identity across a dozen clips without re-uploading images every time.

For a video model with native audio, that reuse is worth more than it is for images. Getting the same face and the same voice across a five-shot sequence is otherwise a manual exercise in tracking down which folder had the good reference. One file, one dropdown entry, done.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/Nougan_Nodes
# restart ComfyUI

Or ComfyUI Manager → "Nougan". RefMod needs a ComfyUI build with MiniMax H3 support; the vendored engine is loaded lazily so a missing H3 module disables only this family, not the whole pack. No pip installs. Press R on the node after baking to refresh the dropdown, and Ctrl+F5 if the character cards render as plain rows.

Where people get burned

"Nothing changed." Either (none) is selected, or all three slots have strengths at 0. The pass-through is silent apart from a console line.

Token anxiety at 4 refs × copies. This is the node where H3's context budget actually bites. copies is the first thing to cut, then slot count, then strength (strength doesn't change token cost at all - it changes influence).

Conditioning from the wrong place. Wire the negative by mistake and the character won't appear; H3's expects positive here.

A face that drifts scene to scene. That's usually reference quality, not the node. Two clear, front-lit portraits beat six stylish ones - the same lesson as every other consistency tool in the ecosystem, and the ones where the face is small or in profile are near-useless.

Licence, again. H3's community licence excludes the US, EU, UK and South Korea from its applicable territory. It's the sort of thing worth knowing before you build a pipeline on it.