MiniMax H3 Build Guide Spec from Video (experimental)

vloMiniMaxH3BuildGuideSpecFromVideo is the "from video" version of Build Guide Spec - and if you've read that one, the shape is familiar. Where the still version plans a single anchored image, this one cuts a masked video into guide clips, aligns each to your target video, encodes them, and weights them, all up front, into one VLO_H3_GUIDE_SPEC. It touches no conditioning. The point is the same as the still-spec node, raised an order of magnitude: when a guide has many frames, several clips, and a bunch of discarding decisions along the way, you do not want two separate implementations of "what counts as a guide" drifting apart. Compute the plan once; let both the latent path and the Qwen semantic path read the same result.

The video-to-clips logic is identical to the direct inline node (Add Masked Guides from Video): frames whose mask is empty are dropped, each surviving contiguous run becomes one guide, and each run is rounded down to an H3-valid length (1, 5, 22, 39, ... frames). The spec then carries all the per-clip weights and anchors with it, and - critically for the semantic side - knows which guides are full-confidence (every token exactly 1.0) and therefore eligible to be shown to Qwen at all.

Inputs

latent (the AV latent these guides are planned against), vae, and video at 24 fps. The anchoring and shaping controls: frame_idx (target frame the guide video's first frame lines up with), strength, min_aug, mask_gamma. Then the segmentation knobs: min_coverage (fraction of the canvas a frame's mask must cover to be worth guiding with), time_pooling (average or max - the union for subjects moving across a latent token's frames), and chunk_align (start or center, which end survives when clip-length rounding drops frames). The mask is optional, one per guide frame, same confidence polarity as the rest of the family - leave it unconnected for a full-confidence guide, which is what semantic conditioning requires.

Outputs

guide_spec - the plan, typed VLO_H3_GUIDE_SPEC, meant for this pack's consumer nodes - plus a plan STRING that spells out what the node decided. With empty-frame dropping, run segmentation and rounding all in play, the plan text is your window into whether the node carved your footage up the way you intended. Feed the spec to Add Guides from Spec (latent half) and/or Apply Semantic Guides (Qwen half); both check the spec's stored geometry against the latent you're actually sampling and refuse a mismatch, because a spec planned on one canvas can't be silently stretched onto another.

Caveats

Same as every node in this family: experimental, needs Patch Masked Guides for masks to mean anything, pinned to recent ComfyUI internals. Install the whole pack (git clone https://github.com/PxTicks/ComfyUI-vlo.git into custom_nodes/, restart, no pip deps), and keep H3's weight/licence reality in mind - ~42.5 GB, with the US/EU/UK/Korea excluded from the community licence.