MiniMax H3 Build Guide Spec (experimental)

An H3 guide reaches the model through two completely separate channels. There's the pixels path: the guide's image is VAE-encoded into condition latent rows that steer the DiT. And there's the semantic path: the same pixels get shown to H3's Qwen encoder as a timestamped <Video k> reference that the text prompt can talk about. Two channels, one observation - and if they disagree about what's in the guide, the model gets told two different things about one moment.

That's the problem vloMiniMaxH3BuildGuideSpec exists to solve. It aligns, encodes and weights a still guide once, into a VLO_H3_GUIDE_SPEC, without touching the conditioning yet. You then feed that one spec to both consumers - MiniMax H3 Add Guides from Spec (the latent half) and/or MiniMax H3 Apply Semantic Guides (the Qwen half) - so the crops, clip lengths and frame indices the two paths use are literally the same object, not two implementations that might drift.

Inputs

• latent - the AV latent this guide is planned against. You should be sampling exactly this latent (the briefs point at taking it from an Empty MiniMax H3 AV Latent and giving the conditioning node the same width/height/length).
• vae - the video VAE.
• image - the single guide image.
• frame_idx - anchor frame; negatives count from the end of the video.
• strength, min_aug, mask_gamma - the weight trio from the masked-guide family: overall scale, what confidence 0 maps to, and the mid-tone curve.
• mask - optional. Guide confidence, same polarity as the rest of the family (1 = trust, 0 = corrupt to noise). Leave it unconnected for a full-confidence guide.

That last point matters more than it looks: Leave unconnected for a full-confidence guide, which is what semantic conditioning needs. Only a guide whose every token sits at exactly 1.0 can be shown to Qwen - a weakened guide is telling the model part of it is unreliable, and Qwen has no way to represent that. So the spec doubles as the honest planner: it knows whether a guide is semantically eligible or not.

The output

One guide_spec output, typed VLO_H3_GUIDE_SPEC - this pack's own custom type, so it won't plug into anything except the pack's consumer nodes. The spec also carries the target's width, height and frame count, and the consumers refuse to apply it to a different geometry: every crop and timestamp was decided on one canvas, so a spec built against the wrong latent is an error, not a resample.

Where it fits and what to watch

This is the newer, spec-based spine of the masked-guide experiment. If you're building a single still guide you can still use the older direct Add Masked Guide, but once you want the Qwen semantic side too, the spec flow is the one that keeps both halves honest. All the family caveats apply: it's experimental, needs Patch Masked Guides on the model for the masks to mean anything, and is pinned to recent ComfyUI internals. Install the whole pack - git clone https://github.com/PxTicks/ComfyUI-vlo.git into custom_nodes/, restart - and remember H3's weights are large and licence-restricted for US/EU/UK/Korea.