H3 SPEED · Save ONE Frozen Stage (T8 EXP)

There is exactly one input. You feed MiniMaxH3SPEEDStageSaveEXPT8 a T8_SPEED_STAGE_RESULT, it writes an immutable artefact and hands back the exact location and digest. That's the whole node.

The value of it is the thing that isn't there: automatic cache lookup. You don't save so that the pack can find your stages later. You save so you can point at one specific stage and say "continue from this". The author's framing - "returns an exact path and SHA for explicit later selection; no automatic cache lookup" - is the entire design philosophy of the modular sampling family in one sentence.

Outputs worth understanding

• stage_result - passed straight through, so you can chain the save in the middle of a graph without rewiring anything downstream.
• artifact_path - a relative manifest path.
• artifact_sha256 - the digest of the written artefact. Both of these go into MiniMaxH3SPEEDStageLoadEXPT8 together. Copy them somewhere; the example resume graphs have empty placeholders in those fields and will not run until you paste in real values.
• report_json - the write report.

Under the hood it's a safetensors file under the pack's own SPEED store in ComfyUI's output directory, and each save creates a new file. Nothing is ever overwritten, which is good for reproducibility and mildly bad for your disk over a long weekend of iteration.

Watch out for the completeness gate. The node rejects an incomplete or unknown sampler execution before it writes a completed artefact - it won't hand you a file and let you discover later that the receipt was a stub. If your save errors out, that's the guard doing its job; read report_json for which receipt check failed rather than retrying and hoping.

What it can, and can't, vouch for

Authenticated external Stage EAV and SPEED Relay calls are supported - the pack can look at those and record real invocation evidence. What it won't do is claim portable restoration for a MODEL, noise provider or effect stack it can't identify. So: if you've got a third-party attention override or an unverified LoRA on a stage, expect the report to say so, and expect a later load to be honest about the limits of what it can re-verify. That's a feature. The alternative is a cache that lies to you three days later.

Install

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

Restart ComfyUI fully and refresh the page - not just reload the workflow. Manager users can search MiniMax H3 Audio T8; if Manager's copy lags the GitHub release, install from GitHub, since the two publish on separate schedules. No Python dependencies to install: the pack deliberately keeps an empty requirements file so it can't replace ComfyUI's Torch/CUDA stack. Recent ComfyUI core is required for native H3.

Gotchas

Cold-resume graphs select by path and SHA. Get one character wrong and the load fails - that's intended, because the alternative is resuming from a different artefact than the one you validated. If you edited the earlier stage's settings after saving, the saved stage is still what it was: the save froze a moment, not a recipe. The pack's docs are explicit that a frozen stage doesn't assert today's upstream settings still match it, so if you changed the prompt on stage 0 and your stage 2 resume looks wrong, it's not the loader's fault.

One housekeeping note: the store grows one file per save per stage. If you're iterating on a three-stage 124-frame clip, you'll be surprised how fast that adds up.