Apply H3 RefMod

Every other node in ComfyUI-MiniMaxH3Mod is building toward this one. Extract makes a reference latent, the loaders stack a few of them into a bundle, and Apply H3 RefMod is where that bundle stops being a file and starts steering the video. It's the only node in the pack that touches the actual conditioning, so if your refs are silently doing nothing, the problem is almost always somewhere on this screen.

What it does

Apply appends each mod's reference latent to the conditioning's minimax_refs list, which is exactly the path a native H3 reference image/video rides - the DiT attends to it through all its blocks like a real ref2va input. The trick is that the pack already compressed that latent for you, so you get reference-style conditioning at a fraction of the token cost.

It's friendly about which conditioning you feed it. The conditioning input accepts either the built-in CONDITIONING type from ComfyUI's core MiniMaxH3ReferenceToVideo node, or MINIMAX_H3_COND from the ComfyUI-MiniMaxH3 pack - it returns whatever type you gave it, so you don't need the sibling pack installed. That migration was the whole point of the unified node: older Apply H3 RefMod (Cond) workflows get rewritten to it automatically.

The inputs that matter

• retention - your master strength, multiplied with each loader row's strength. The tooltip maps it to MiniMax's levels: 1.0 = fully_preserved, 0.7 = partially_preserved, 0.4 = attribute_transfer (style and attributes, not identity), 0.15 = weak_reference, 0 = nothing. Start at 1.0 and only drop it when the ref is drowning the scene.
• curve_direction - the one that confuses everyone. It shapes an envelope across the ref's own frames (stacked images or video-ref latent frames), deciding which reference content dominates. It does not schedule anything in the output video. constant (the default) keeps every ref frame at full strength; the old concept_at_end default faded early frames toward blur, which roughly halves average strength on multi-frame mods. Old saved workflows keep their old values, so if you load a shared graph and it feels weak, check this first.
• override - flip this on and Apply reads retention + curve from config baked into the mod by Fix H3 RefMod Config instead of the widgets. If no mod in the bundle carries a config it prints a note and falls back to the manual params, so it never silently does nothing.
• scramble_seed - -1 (default) keeps refs in order. Give it a seed and it shuffles the bundle (and can keep a random subset), so a multi-ref mod "pops" a different character per run. Set its control-after-generate to randomize for per-run variation.

The curve graph output is an optional 1024×1024 debug image of the envelope - useful when you're using presets or override and want proof of which curve actually ran. Leave it unconnected.

Install

ComfyUI Manager, search "ComfyUI-MiniMaxH3Mod", or:

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

Restart ComfyUI. Python deps (safetensors, numpy, Pillow) install automatically; nothing heavy. The pack downloads no models - but Apply is useless until you have real MiniMax H3 weights plus its video VAE in ComfyUI, and remember the H3 Community License excludes the US, EU, UK and Korea from running the local weights.

Common issues

Where people get burned: forgetting that Apply strength is a product - loader strength 0.3 times retention 0.4 is 0.12, which is basically air. Watch the console: Apply prints how many ref blocks were injected and the retention used. And don't reach for curve_direction when you want a ref that fades in over the generation - that's what H3 RefMod Step Curve is for, since it runs over the denoise timeline instead of the ref stack. The pack's README is upfront that neither axis is a guaranteed "identity then style" split; think of curves as levers, not promises.