H3 HyperFlow P7 · Load Exact Completed HIGH (T8 EXP)

You spent an hour getting a segment to HIGH. The run finished, you saved it, and then you realise you want to re-encode the candidate at a different CRF, or just re-mux it, or hand the same finished AV into a delivery graph tomorrow. You do not want to sample it again. This node is that: it reads a completed HIGH artifact off disk by exact path and SHA and reconstructs the typed result, so everything downstream - candidate save, decode, delivery - works as if the sampler had just run.

How it works, and what it is not

It is a selected artifact load, not a cache hit. The distinction matters. ComfyUI's normal caching is about "nothing upstream changed, skip it"; this is about "I have a file, its bytes hash to what I expect, and I explicitly want to reuse it." There's no automatic matching of today's widget values against the frozen file's settings. If you edited the HIGH model, prompt, or noise since saving, loading the old artifact does not tell you that, and the pack is upfront about it.

What it does do is revalidate the chain: it takes the live high_handoff and checks the current P7 parent, the LOW lift, the HIGH conditions and the reconciled handoff against what the artifact recorded. So a changed parent or a rebuilt lift fails the load instead of quietly producing a segment that doesn't belong to this chain.

Inputs and outputs

high_handoff - yes, it needs the live handoff even though it isn't sampling. That's the whole point of the revalidation; you can't load a HIGH result into a void. artifact_path and artifact_sha256 come from Save Exact HIGH 4:8, which returns both. The path is relative to the chain's stage store; the SHA is the integrity check. Copy them, don't retype them.

Outputs mirror the fresh path exactly: high_result (typed, for candidate save or another save), completed_av (the LATENT, if you're decoding it yourself), and report_json. Downstream nodes cannot tell the difference between this and Bind Completed HIGH Result - which is the design goal.

Getting the path and SHA

The save node writes under your ComfyUI output directory, in the chain's own P7 namespace:

output/MiniMaxH3/modular_p7_stages/<chain_id>/segment_00000/...

Both the path and the SHA print in the save node's report_json and are returned as sockets. Wire them, or paste them, but never guess. The example workflows ship placeholder values here - including a 64-zeros SHA - and loading with a placeholder errors out rather than falling back to anything helpful.

Install

Manager → MiniMax H3 Audio T8, or the usual clone:

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

Then fully restart ComfyUI and refresh the browser. The pack's requirements.txt adds no packages on purpose, so this install can't brick a working CUDA setup. The load itself needs no model files - that's rather the point - but the rest of the graph around it still needs the H3 base model, both VAEs, the HyperFlow adapter in models/hyperflow/loras/, and the learned 3D latent upscaler in models/latent_upscale_models.

Failure modes

Missing file, wrong SHA, a path that escapes the chain directory, a partial write, a type mismatch or a tensor that doesn't match the manifest - all of those fail closed with an explicit error. That's good, but it means "my path is right, why is it rejecting" is usually one of:

• the artifact was written by an older chain with different models/prompts, so revalidation is correctly refusing it;
• you copied the path but let your shell or file manager hand you a resolved absolute path with a different case;
• the file is still being written, or another process holds the lease.

Last thing worth knowing: if HIGH used Prompt Relay, resume isn't a free decode - rebuilding the HIGH source can still pull in the HIGH model. The example workflow README flags this explicitly, and it surprises people who assumed "loaded" meant "no model load".