H3 HyperFlow · Load Exact Frozen HEAD (T8 EXP)

Two inputs, two outputs, and it replaces half your graph.

You give it an artifact path and an exact SHA-256 - both mandatory, both defaulting to empty strings - and it gives you back a continuous_boundary that plugs straight into a continuous TAIL stage. The good news is what's now missing from your workflow: the HEAD model, its conditioning, its noise, its effects and its sampler. The pack's own resume examples genuinely remove those nodes rather than leaving them muted, which is the difference between a lighter graph and a lighter graph you're pretending is lighter.

It loads nothing else. No diffusion model, no CLIP, no sampler.

What's still checked

Not everything is a formality after a load. When the boundary reaches the TAIL sampler, the pack verifies that the same original base and the same HyperFlow adapter contents are in play - because the artifact is a state captured under a specific two-time contract, and injecting it into a different-looking patched model would produce something that runs and means nothing.

Practically, that means the TAIL half keeps its own freedom - independent content LoRAs, independent prompt, independent effects, and even a different model branch if it's built from the same full H3 base through the dedicated HyperFlow loader - while the base identity stays pinned.

Failures are fail-closed. A missing file, a wrong SHA, a corrupt artifact, a path that escapes the storage root, a lock held by another process, or a contract mismatch all error out. The node will not look for a similar file, and it will not silently regenerate the HEAD half. The docs put it as: missing/corrupt/busy artifacts fail, no hidden regeneration. Which is correct - a silent 8-step re-run you didn't ask for is worse than an error.

Inputs and outputs

In: artifact_path, artifact_sha256.

Out: continuous_boundary and report_json. The boundary is deliberately the same type the HEAD stage produces, so a downstream TAIL stage, a HEAD/Tail audit or another save node can't tell the difference - the type carries the actual interval, and the TAIL stage reads the split point from it rather than from a widget you might have changed.

That's also why you shouldn't hand-edit anything in the artifact. Renaming the file is fine if your path widget follows; touching the tensors or the manifest turns a working resume into a checksum failure. And do keep the hyperflow_stage_artifacts directory out of any "clean up old outputs" script.

The mental model, one sentence

Loading a frozen HEAD is an explicit freeze of a specific executed state, not ComfyUI deciding your inputs look unchanged and skipping work. There's no recipe cache and no equivalence claim. If you've edited the HEAD prompt or swapped its content LoRA since saving, the artifact still describes the old run; the pack won't pretend otherwise, and neither should your expectations.

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 afterwards, recent ComfyUI core required, and for the TAIL half you'll still need the H3 base model, Qwen3-VL encoder, both VAEs and the original HyperFlow adapter in models/hyperflow/loras/. The pack's empty requirements.txt is deliberate: no pip step, no risk of swapping ComfyUI's Torch/CUDA build.

Where people get burned

Placeholders. The example graphs ship with zeroed 64-bit SHAs and blank paths, and they will not run until you fill them in from a real save.

Mismatched runs. Resuming a HEAD artifact from a different base model or a differently-patched HyperFlow setup is exactly the case the contract check refuses. That usually means you changed something legitimate - a content LoRA on the HEAD branch - in which case the honest fix is re-running HEAD and saving again, not trying to force the resume.

And patience with the report. It's the only place that tells you which artifact you actually loaded and what the TAIL stage is about to inherit. Thirty seconds of reading beats a twenty-minute render of a mismatched half.