H3 Progressive · Save Frozen LOW Boundary (T8 EXP)

Workflow iteration on H3 has a shape problem: the expensive part and the part you're fiddling with are in the same graph. In a two-stage run, once LOW is done, everything you actually want to tweak - the HIGH prompt, the HIGH LoRA, the effects on the full-size pass - is downstream of it. Re-running LOW every time you change a word is how an afternoon disappears.

This node writes the LOW boundary to disk as a typed artifact with an exact hash. Then the load node gives you a HIGH-only graph that starts from those bytes.

What it does

In: low_boundary and a prefix. Out: the same low_boundary passed through, artifact_path, artifact_sha256, and report_json. It's an output node, so it writes files as part of the queue.

Three things about the file it writes:

• It lands under output/MiniMaxH3/progressive_stage_artifacts, keeping stage artifacts away from your normal output folder. The path is returned so you can copy it straight into a load node.
• It's a new artifact, never an overwrite. The docs are explicit about this: no existing result gets replaced, so you can't destroy yesterday's good boundary by running today's experiment.
• It's not an ordinary stage save. This artifact keeps evolving audio and clean video distinct, because the boundary isn't a finished clip - the audio stream is deliberately mid-flight. A generic "save latent" node would collapse exactly the distinction HIGH depends on.

An honest caveat that comes with it: unknown executable identity may be archived but not certified for persistent reuse. If you've wrapped the model in third-party patches the pack can't authenticate, the save still works - it just won't claim your artifact is safe to reuse forever.

Practice

... LOW Sampler Only ─→ Save Frozen LOW Boundary ─→ copy artifact_path + artifact_sha256 into a text note

Then build the cold graph: Load Exact Frozen LOW Boundary → Boundary to External Lift Input → upscaler → HIGH Handoff Only → HIGH Sampler Only. That graph has no LOW model, no LOW conditioning and no LOW sampler in it at all, which is the real payoff - it runs in a fraction of the memory, and it can't accidentally re-sample LOW.

Save both values. The SHA is not decoration: it's what the load node selects on, and a path without a hash is a path that could be anything.

Install

Manager, search MiniMax H3 Audio T8. Or, from a terminal:

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

Restart ComfyUI fully, then refresh the browser page. The pack ships no installable dependencies - its requirements.txt deliberately won't touch Torch/CUDA - so if the node doesn't appear, the process wasn't actually restarted.

Disk space is the one resource this node really uses. Boundaries are not small, and there's no overwrite, so old experiments accumulate under progressive_stage_artifacts. Prune them by hand now and then.

Where people get burned

Assuming a saved boundary proves today's settings match. It doesn't, and the pack says so twice on purpose. Change the LOW model, LoRA, first frame, prompt, canvas, plan or effects and the frozen boundary is stale - re-run LOW and save a new one. The SHA proves bytes, not intent.

Reusing a prefix and expecting a tidy file. You'll get a second unique artifact, because no-overwrite is the contract. That's safer than it sounds and occasionally confusing when you're looking for "the" boundary in a folder of five.

And copying the pair by hand into the load node both times. If you only fill in the path and leave the hash blank, the load is a request you didn't mean to make. Fill both, every time; it takes five seconds and it's the difference between resuming a run and resuming your run.