H3 HyperFlow · Save Frozen HEAD (T8 EXP)

The continuous HyperFlow route hands the second half an exact model-space x_sigma state captured mid-trajectory. If you want to iterate on the TAIL half - different prompt, different content LoRA, different effects - you should not have to re-run the first half every time. This node writes that raw state to disk so you don't.

The important word is raw. What gets saved is the model-space x_sigma plus the separate Core scaffold, not a completed clean x0. There's a sibling Tail Save node for finished AV results; this one is mid-trajectory state, and the two are not interchangeable. Try to feed a HEAD artifact into an upscaler and you'll find out why the distinction is worth spelling out.

Inputs and outputs

Two inputs. continuous_boundary - straight off a HEAD stage or off the HEAD effects audit, either works - and prefix, a string defaulting to HEAD that names the artifact. Set it to something meaningful if you're generating more than one; you will be comparing these files later and HEAD tells you nothing about which take it was.

Out: continuous_boundary (passed through, so the chain keeps going), artifact_path, artifact_sha256 and report_json. This is an output node, so it can terminate a branch. Copy both the path and the SHA - the load node wants both, and the SHA is what turns "a file that looks right" into evidence.

Artifacts land under output/MiniMaxH3/hyperflow_stage_artifacts. Each save gets a new directory: nothing is overwritten, the write is protected by an OS lock, and the manifest is committed last, so a crash mid-save leaves you with a rejected artifact rather than a corrupt one that silently poisons a resume. Safetensors, no pickle.

The honest caveat about ownership

This is where the pack gets unusually careful, and it's worth understanding rather than skimming.

The artifact binds the frozen boundary to the contract that will interpret it: the actual HyperFlow closure, the delegate, endpoint weights, FP32 endpoint caches, the original weights, and the selected patch contents. That's a lot more than a filename. If your model stack includes unknown user patches, the pack will still execute and archive them - but it won't hand out certified cross-process reuse for them. You may be able to reuse the artifact in the same process; a fresh process with modified patches may refuse it, and that refusal is the design working.

Translation: the save/load pair is reliable when you're running the pack's own HyperFlow path with content LoRAs it recognises, and best-effort when you've stacked third-party attention hijacks on top. Don't build a workflow that silently depends on the latter.

Install

Manager → 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. Needs a recent ComfyUI core, the H3 base model, Qwen3-VL encoder, video and audio VAEs, plus the original non-pruned HyperFlow adapter in models/hyperflow/loras/ - no automatic download, and a pruned base is rejected rather than approximated. No pip packages come from this pack by design.

Where people get burned

Losing the SHA. It's returned as a string; if you close the tab without copying it, you'll be hunting through manifest JSON files by hand. Paste it straight into your load node (or a note node) before you queue the next run.

Expecting resumption to be free. A frozen HEAD means the TAIL half still runs: model, content LoRAs, conditions, effects, and its execution seed - which, note, is a model execution seed, not a second noise seed. There's no new noise on this route at all; the whole point of the continuous contract is that the second half picks up the exact state the first half left behind.

And the description trap, if you're reading node descriptions rather than docs: some of the older HyperFlow node descriptions still say disk resume is pending. Those lines are stale by the author's own admission - old schemas were frozen, and the save/load entries exist now. The modular sampling documentation supersedes them.