H3 HyperFlow P7 · Bind Completed HIGH Result (T8 EXP)

Kinda boring node, genuinely important one. Bind Completed HIGH Result takes the sampler's receipt and the handoff that produced it, checks that the fresh 4:8 stage actually completed and that the source tensor is the one this phase was set up for, and hands you a typed result. Nothing samples, nothing decodes. It's a verification step with a socket.

You might reasonably ask why it exists instead of just wiring the sampler's latent output straight into the candidate save. The answer is that in the P7 route a segment is a provenance record, not just a picture: the pack wants every stage transition to be evidenced by an immutable receipt rather than inferred from a wire. That's why the effect audits exist too - "an Apply node alone is not proof of execution" is the same idea in a different place.

Inputs

high_handoff - the typed output of Original HIGH Audio Reconcile. This is what lets the bind node check the result against the phase that requested it.

stage_result - the T8_STAGE_RESULT from your stage sampler. In practice that's the ONE Stage Sampler node the pack ships for this purpose. A plain SamplerCustomAdvanced gives you latents, not a stage result, and this node won't accept it.

Two inputs, no widgets. If you were hoping for a knob here, the knob is upstream.

Outputs

high_result is the thing you actually wire - into Save Exact HIGH 4:8, into Save Separate Candidate, or into both. completed_av is the LATENT; it's the terminal joint AV for the segment, so if you're rolling your own delivery you can decode it directly. report_json carries the verify receipt.

Note what the receipt is not: it is not a quality check. The pack is unusually careful about this wording, and it's right to be. "The stage completed and the tensor is the one we asked for" and "this segment looks good" are different claims, and only you can make the second one. That gap is also why the candidate node doesn't auto-accept anything.

The bigger point of the route

P7 splits a long video into 124-frame segments with a 22-frame context overlap, runs LOW 0:4 → learned 3D lift → HIGH 4:8 per segment, and chains segments through explicitly accepted candidates. The payoff is that a second segment can be re-run in isolation: change the HIGH prompt or the HIGH LoRA, sample four steps, and you're not paying for the LOW pass again. The cost is bookkeeping - chain ids, job SHAs, manifests. Nodes like this one are the bookkeeping.

Worth knowing before you commit: H3 itself is a 33B omni-modal model with joint audio, so the memory floor is not small, and P7 runs two phases plus a latent upscaler. A 4+4-vs-8-step comparison is also not a free upgrade - step-reduced pipelines always give something up, which is why the pack refuses to certify image quality on your behalf.

Install

ComfyUI Manager → search MiniMax H3 Audio T8, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Full restart, then browser refresh. The pack ships an intentionally empty requirements.txt - no pip installs, so it can't overwrite the Torch/CUDA stack ComfyUI is built on. You still supply H3 weights (models/diffusion_models, models/text_encoders, models/vae), the HyperFlow adapter under models/hyperflow/loras/ (the original, non-pruned file - the loader rejects pruned ones), and a learned 3D upscaler in models/latent_upscale_models.

If it errors

"Stage result doesn't match the phase" almost always means the receipt came from a different phase or a stale run - re-queue the graph rather than re-using a cached stage result. If you're seeing this node pinned red before anything samples, the usual cause is that stage_result is being fed by a core sampler instead of the pack's stage sampler; the types are deliberately not interchangeable.