H3 HyperFlow P7 · Save Exact LOW 0:4 (T8 EXP)

LOW is the first of P7's two passes: four steps at a fraction of the target resolution. It's the cheaper half by a wide margin, and once it's done, the content of the segment - the composition, the blocking, the movement - is more or less decided. What you'll iterate on afterwards is the HIGH pass.

So save the LOW. That's this node: it writes the completed LOW result as an immutable artifact, hands back the path and the hash, and gets out of your way.

What gets written

A new artifact in the chain's own P7 stage store, alongside the HIGH artifacts but a distinct type:

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

The P7 store is separate from older HyperFlow caches and from the generic stage store on purpose - nothing overlaps, so nothing can be silently mistaken for something else. Saves are atomic, one directory per save, OS-locked, no pickles, and the manifest is committed last so a partial write is detectable rather than loadable.

The artifact holds the completed LOW result with its sampler receipt and the exact provenance: which accepted parent, which phase, which source tensor. That's what makes it loadable later by Load Exact LOW for HIGH Only.

Inputs and outputs

One input: low_result, from Bind Completed LOW Result (or from a LOW load, if you're re-saving a reconstructed stage).

Outputs are the usual pass-through pair plus the two values you must keep: low_result and low_x0 continue downstream, while artifact_path and artifact_sha256 are what you feed the load node. report_json is the manifest.

The path is relative to the chain directory and the hash is mandatory on the load side, so treat the pair as one value - copy them together into a text node or into your notes. The example workflows literally ship a 64-zero placeholder hash that fails until replaced.

The honest scope

Saving freezes the result, not the settings. Reopen this artifact later and the loader will check what it can - the current accepted parent, the freshly prepared LOW phase - and reject if the chain has moved underneath it. It will not tell you that you've since changed the LOW model, prompt, noise or effects, because the format doesn't claim to record today's graph, only yesterday's output. The author's description says it outright: a frozen selection is not an automatic cache hit.

Fair warning on portability too: artifacts saved from an executable stack the pack can't identify can be archived, but they don't get certified cross-process reuse. If you've stacked third-party patches onto the LOW model, expect the loader to be conservative.

Install

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 of ComfyUI and a browser refresh. The pack's requirements.txt is intentionally empty - no extra packages, so installing this can't swap out the Torch/CUDA build ComfyUI depends on. You still need the H3 base model, Qwen text encoder, both VAEs, the original non-pruned HyperFlow adapter in models/hyperflow/loras/ and the learned 3D latent upscaler in models/latent_upscale_models.

Housekeeping

P7 is disk-hungry: two stage artifacts per segment, plus candidate video files in a separate candidates/ tree, plus whatever your ComfyUI install keeps in its temp folder. The stage artifacts are the small, useful ones; the candidates are the big, disposable ones. If you're cleaning up, delete rejected candidate directories and keep the stages you might resume from.

One chain id, one project. The store path is keyed on <chain_id>, so two runs sharing an id share a directory - and the chain lease means you can't run them at once anyway. Boring, but it's the kind of thing that costs an afternoon when you're six segments deep and can't remember which id was which.