MiniMax H3 Visual Reference Strength (EXP/T8)

If you've generated image-to-video or Ref2VA on MiniMax H3 and thought "this looks like the model sanded my reference smooth," this is the node you were looking for. H3 holds onto your visual reference hard - that's great for identity, but it also means reference/keyframe latents get baked in at near-full strength and fine texture can come out plasticky. MiniMaxH3VisualReferenceStrengthEXPT8 is one knob that turns that down.

How it works

H3's conditioning carries a parameter called minimax_visual_cond_noise_aug. The author's tooltip is blunt about it: this node maps your reference_strength value directly onto that parameter. When you lower the value, the node adds seeded noise to the visual reference/keyframe latents before they reach the sampler. The model then sees a slightly degraded reference and - this is the theory - has to fill in detail itself instead of copying the smoothed version, which can reduce that over-polished look. The tradeoff is exactly what you'd fear: go too low and identity and composition start drifting, because the reference isn't anchoring the output anymore.

The inputs are just two

positive is the conditioning you want modified (must be a Conditioning that actually contains H3 image/video refs or keyframes - the node can't add reference strength to a prompt that has no reference). reference_strength defaults to 0.999 and the tooltip gives you the exact ladder the author recommends: try 0.999, then 0.995, then 0.990, and treat anything ≤0.950 as aggressive. That's a tight window because it's a global modifier on an already-strong conditioning path - small nudges, then look, don't swing it.

Outputs:

positive (the modified conditioning, which you wire onward into the sampler) and a report string with what it did. In-out on the same socket, so you drop it inline between your conditioning and your sampler. It never loads a model and adds effectively zero cost.

Installing:

part of the MiniMax H3 Audio T8 pack. ComfyUI Manager → search "MiniMax H3 Audio T8", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart after. No extra pip dependencies, no weights to fetch.

Where it bites:

this is a global setting, so it hits every visual reference in the conditioning, not just the one bothering you. If you have a strong identity shot and a weak texture problem on a secondary element, a global nudge might buy texture and lose the face. The other gotcha is that this is the experimental end of the pack - the author's own framing is "may reduce overly smoothed texture, but can weaken identity and composition," so treat it as a test knob, not a set-and-forget. And remember the model-level context: H3's local weights are territory-restricted (no US/EU/UK/Korea under the Community License), which gates all of this regardless of how the knob feels.