H3 Condition Strength

This node is a single dial for a question H3 forces you to answer: when you feed it a keyframe or a reference image, how much should it actually trust it? H3ConditionStrength takes a CONDITIONING, applies the two strengths, and hands the same conditioning back. And the first thing to know is that its own description is the honest one - it does nothing on a plain text prompt. If your graph has no keyframes and no references, this node is a pass-through. It only matters once you're anchoring.

How it works

The mechanism is noise, used deliberately. Under the hood, keyframe and reference images get encoded to latents and injected into the conditioning. visual_strength controls how clean those image latents are fed in. At 1.0, they're pristine - strongest adherence, the model locks onto your supplied frame. Below that, the node blends seeded noise into them, and noise gives the model permission to drift: the frame becomes a suggestion instead of a law. Default is 0.999, which is the shipped recipe's way of saying "essentially pristine but with a hair of slack so the model can actually render motion instead of freezing on your still."

audio_strength is the same idea for reference audio latents - ref2va only, since fl2va has no reference rows and ignores audio refs entirely. Default 1.0. If a voice anchor is coming through too strongly, this is where you loosen it.

When to touch it

In the keyframe workflow, it's your main lever when a clip is either glued to an anchor frame (motionless, literally matching the still) or ignoring it entirely (drifted off into a fresh interpretation). Turn the strength down from 0.999 to, say, 0.95 and the model gets room to actually move while still holding composition. Turn it toward 1.0 when you're anchoring identity and you need the face to not wander.

One genuinely useful pattern from the pack's own tooling: per-anchor strength. The Keyframes workflow uses this node to give different anchors different authority - a start frame you want obeyed hard, a mid anchor you're willing to let breathe. Same node, different values per lane.

Install

It ships in the H3 Multishot pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

or ComfyUI-Manager → H3 Multishot. Restart; ComfyUI v0.30.0+. No dependencies to install beyond the pack itself.

The trap to remember: this is a post knob, not a prompt fix. If your ref is being ignored entirely, check the loader pairing (a GGUF encoder without its vision sidecar will silently drop image refs - see the H3 CLIP Loader page) before you start tuning strength. The dial only adjusts how hard existing conditioning bites; it can't put back conditioning that never arrived.