MiniMax H3 Add Masked Guide (experimental)

MiniMax H3's stock Add Guide lets you drop a still image into a video at some frame, full stop - the whole image steers the whole generation at one strength. vloMiniMaxH3AddMaskedGuide is that node with a spatial opinion: you anchor an image at any frame of the video, and a mask says which parts of that image to trust and which to treat as unreliable. Want a character's face locked in but its background free to change? Mask out everything but the face. That's the use case.

The name is doing double duty, so read it carefully: the mask is guide confidence, not a denoise mask. In a denoise mask, 1 means "generate here". Here it's flipped on purpose - 1 keeps the guide at full strength, 0 corrupts that region of the guide to noise, and values in between blend continuously. The pack made this deliberate inversion so the node reads as a trust map, not an inpaint map. If a mask that "should" keep the image is zeroing it out, you've got the polarity backwards in your head.

How it works

The guide's pixels are VAE-encoded into condition tokens (H3's 1×2×2 patch grid). This node pools your mask onto that grid by area averaging, so a token half covered by the mask lands at 0.5 and soft edges survive, then attaches a per-token strength to the guide's keyframe. On its own that does nothing - ComfyUI's stock forward pass ignores the extra data. That's the job of MiniMax H3 Patch Masked Guides, which has to sit on your model somewhere in the graph. Patch present: each token's strength becomes its own noise-corruption level and its own condition timestep. Patch absent: the mask is ignored and the guide behaves exactly like a stock Add Guide. Fully open mask (1 everywhere, strength 1.0): bit-identical to stock.

Inputs you'll actually set

• positive / latent / vae / image - the conditioning you're extending, the AV latent you're sampling, your video VAE, and the single guide image. (Guide clips aren't supported here yet - a still is the unit. For masked video guidance, use the "from Video" sibling.)
• mask - the confidence mask, framed on the same crop as the image.
• frame_idx - where to anchor it. 0 is the first frame; negative values count back from the end, so -1 pins the guide to the final frame.
• strength - scales the whole mask (default 1.0).
• min_aug - what a mask value of 0 maps to: 0.0 replaces those tokens with pure noise; raise it to keep a floor of guidance everywhere.
• mask_gamma - shapes mid-tones: >1 pushes them toward weak guidance, <1 toward strong.

Output is the extended positive CONDITIONING, straight into your H3 sampler.

The usual caveats

This whole family is experimental research code, and the patch node is pinned to specific ComfyUI internals (0.33.0, with a source-fingerprint check) - the masked-guide articles aren't a beginner's first H3 workflow, and the README says so. Mask alignment is where most time goes; that's exactly what the pack's Guide Token Mask Preview node exists to debug. Install is the whole pack, git clone https://github.com/PxTicks/ComfyUI-vlo.git into custom_nodes/, restart, no pip deps - and remember H3's weights are ~42.5 GB under a licence that excludes the US, EU, UK and Korea.