H3 Stage Result · Save Frozen Artifact (T8 EXP)

MiniMaxH3StageSaveEXPT8 takes a stage_result and a prefix, writes an immutable artefact under ComfyUI's output directory, and gives you back the exact path and SHA-256 to get it again later.

Two inputs. Five outputs. The interesting part is the gate in the middle.

The gate

Saving is not "dump whatever latent is lying around". The node refuses to write a completed artefact for an incomplete or unknown sampler execution - the receipt has to say the recipe completed. If you've got a custom sampler from another pack in the chain, or a run that aborted halfway, you'll get an error instead of a file. That's deliberate: a save is a promise to your future self, and the pack won't let you promise something it can't stand behind.

So the first thing to do when a save fails is read report_json. It reports why the receipt didn't qualify. Retrying and hoping is not a strategy here.

Outputs

• output and denoised_output - the stage's latents passed straight through, so the save can sit in the middle of a graph without forcing a rewire. This is the same pair MiniMaxH3StageLoadEXPT8 returns, and keeping them distinct matters: denoised_output is what the learned 3D upscale handoff wants when you resume only the HIGH stage, while output is the terminal state other adapters consume.
• artifact_path - a relative manifest path.
• artifact_sha256 - the digest. Both of these are required inputs on the load node, so copy them somewhere durable; the cold-resume example graphs ship with empty placeholders that you have to replace before queueing.
• report_json, plus the path and digest echoed as text in the UI panel.

Artefacts go under output/MiniMaxH3/stage_artifacts, one new file per save. Nothing is overwritten, which is exactly right for reproducibility and exactly wrong for your free disk space if you iterate a lot.

What it will and won't certify

Authenticated external stage effects - the EAV apply and the Relay adapters - are recorded so a later load can re-verify them. What the pack will not do is claim portable restoration for a MODEL, a noise provider or an effect stack it can't identify. Third-party attention overrides and unverified LoRAs land in the report as non-portable, and a resume from that artefact will be honest about the limits of what it can confirm.

This is the same philosophy running through the whole family, and it's the reason these nodes feel fussy: a cache that says "yes, same result" when it isn't is worse than no cache. If you've ever had a workflow serve you a stale latent and spent an hour thinking your prompt stopped working, you know the feeling.

Install

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

Fully exit ComfyUI and restart it, then refresh the page in your browser. Manager users can search MiniMax H3 Audio T8; when the registry copy lags the GitHub release, install from GitHub - the two publish independently. There's nothing to pip install: the pack keeps its requirements file empty on purpose so it can't replace ComfyUI's Torch/CUDA build. You do need a recent ComfyUI core, because the pack relies on native MiniMax H3 support and the newer node API.

Gotchas

The save is a checkpoint of this run. It is not an assertion that today's LOW settings still match it - the docs make that point directly, and it's the mistake that makes people think their resume is broken when actually they edited the upstream stage after saving.

Pairs of these turn up in examples/workflows/40-modular-two-pass/ (S01/S02/S03 native splits) and 41-vdn-two-pass/: a Full_Stages graph that runs and saves both stages, and a Cold_HIGH graph that loads the saved LOW and runs only the later work. Start with the full one, record the path and SHA, then paste them into the cold one. Never guess a SHA - the load compares it and will reject you.