MiniMax H3 Add Masked Guides from Video (experimental)

A single still guide gets you one anchored image inside an H3 video. Sometimes you want to steer with footage - a clip of the subject in motion, or several distinct shots - and trust different parts of it at different times. vloMiniMaxH3AddMaskedGuidesFromVideo is the node that cuts a masked video into guide clips and anchors each one at the right place in your target video. It's the segmentation case this experimental family was really built for.

The flow: you supply guide frames (at your target video's frame rate, one mask per frame), and the node walks them. Frames whose mask is empty guide nothing and get dropped; each surviving contiguous run becomes one guide clip; and each run is rounded down to a length H3 actually accepts - the model's valid clip lengths are 1, 5, 22, 39, ... frames, not arbitrary counts, so a 7-frame run becomes a 5-frame guide. The mask means confidence again, not denoise: 1 keeps that part of the guide at full strength, 0 corrupts it to noise, in-between blends continuously, and mask_gamma/strength/min_aug shape that as in the rest of the family.

Inputs worth understanding

• positive, latent, vae, video, mask - conditioning, target AV latent, video VAE, guide frames, and one mask per guide frame (a single mask is applied to all of them).
• frame_idx - the target frame your guide video's first frame lines up with. Negative counts from the end. Guide frames that fall outside the target video are dropped.
• min_coverage - a frame is dropped when its mask covers no more than this fraction of the canvas (measured after fitting the guide to the target's framing). 0.0 drops exactly the empty-mask frames; raise it to also drop frames where the subject is barely visible.
• time_pooling - how the masks of the frames behind one latent token combine: average matches the spatial pooling, max takes their union - which is what a subject moving across those frames needs.
• chunk_align - when rounding a run down to a valid clip length drops frames, start keeps the head of the run, center keeps its middle.

Outputs

positive - your conditioning extended with the masked guide clips - and a plan STRING describing what the node decided to do. The plan output is genuinely useful for the video case: with frame-dropping, clip-length rounding and anchoring all in play, you want to see which runs became guides and where they landed before you blame the model for a weird result.

Standing caveats

Two things gate all of this: the model needs Patch Masked Guides on it or the masks are ignored (clips behave like stock guides), and the patch is pinned to specific ComfyUI MiniMax-H3 internals with a source-fingerprint compatibility check - experimental code, new pack, expect to update ComfyUI or pin it. There's also the direct-node versus spec question: this node does everything inline; the newer Build Guide Spec from Video flow plans the same operation as a reusable spec so the Qwen semantic half can read it too. Install: clone https://github.com/PxTicks/ComfyUI-vlo.git into custom_nodes/, restart, no pip deps.