MiniMax H3 Add Guides from Spec (experimental)

If the masked-guide family splits every guide into a latent half and a semantic (Qwen) half, vloMiniMaxH3AddGuidesFromSpec is the latent half. You hand it a guide_spec - the plan produced by Build Guide Spec or Build Guide Spec from Video, where the images were already aligned, encoded and weighted - and it anchors the spec's clips onto your conditioning as masked guides. All the heavy lifting happened upstream; this node is where the plan becomes part of the actual sample.

The schema tells the story. It takes exactly three things: positive CONDITIONING, the latent being sampled, and the guide_spec. Notice what's not there - no VAE, no image, no mask, no frame index. The spec already did all of that against a specific latent, so this node only needs to know which latent you're sampling to confirm the plan still fits, and then extend your conditioning. Output is positive again, one hop further toward your H3 sampler.

Why the spec pattern beats doing it inline

Both conditioning paths need to describe the same pixels at the same moment. If you aligned a guide twice - once for the latent path, once for the Qwen semantic path - every crop decision, clip-length rounding and frame index would have to be reproduced exactly by two separate implementations, forever. A spec computes the plan once and both consumers read it. Add Guides from Spec is where the latent side of that contract is fulfilled.

There's a real safety net hiding in that latent input: the spec remembers the width, height and frame count it was planned against, and this node refuses to apply it to anything else. Since the spec is typically built before the conditioning node creates its own latent, the two geometries get stated twice - and a mismatch is not something you can patch over, because every crop and timestamp in the plan was decided on the other canvas. The error message tells you exactly what the spec expected and what you gave it.

The workflow shape

For a still guide, the chain looks like: Empty MiniMax H3 AV Latent → Build Guide Spec (with your image, mask if any, frame index) → Add Guides from Spec feeding the same latent and the spec → sampler. The spec also flows to Apply Semantic Guides if you want the guide visible to Qwen - but remember only full-confidence guides (no mask, strength 1.0) qualify for that side, and masks do nothing at all unless Patch Masked Guides is on your model. Samples with no masked guide take the stock path, so leaving the patch in place is safe.

The honest caveats

This is research code from a young pack, pinned to specific ComfyUI MiniMax-H3 internals; if the patch node's compatibility gate complains, your ComfyUI build has diverged from what was tested. Install is the whole pack - Manager search "ComfyUI-vlo" or git clone https://github.com/PxTicks/ComfyUI-vlo.git into custom_nodes/, then restart. H3 itself: ~42.5 GB of weights under a licence that excludes the US, EU, UK and Korea.