H3 Progressive · Save Completed HIGH AV (T8 EXP)

You've run the second pass. It's good. And then, like everyone who has ever generated anything, you close the tab - and the only copy of your result is a decoded video, which is the version you can't iterate on any more.

This node writes the completed HIGH AV latent to disk as a typed artifact with its execution receipt, exact output identity and metadata. Next time, you can load it and decode without touching a sampler.

What it does

In: high_result (from HIGH Sampler Only), and a prefix. Out: av_latent (passed through, so you can keep decoding in the same run), artifact_path, artifact_sha256, report_json. It's an output node, so the write happens as part of the queue.

Two contracts worth stating plainly, both from the node's own description:

• It writes a unique new artifact and never overwrites an old result. Running an experiment doesn't destroy the previous run's output - which, for a file you might need in a week, is the right default even if it means pruning folders later.
• It performs no additional sampling or learned lift. It's a save. If the numbers in your artifact differ from what the sampler produced, something else in the graph did it, not this node.

The difference from Save Frozen LOW Boundary is the shape of what's saved. LOW saves a mid-trajectory state - clean video, evolving audio, original audio noise - because HIGH still has to continue from it. HIGH saves a completed joint AV result with its receipt, because nothing downstream needs to continue anything. Different types, different consumers.

Wiring

HIGH Sampler Only ─→ high_result ─→ Save Completed HIGH AV ─→ av_latent ─→ your decode path
                                          └→ artifact_path + artifact_sha256 (copy these)

Both outputs get used: the latent for decoding in this run, the path and hash for Load Exact Completed HIGH AV in a future one. Copy both values the moment the queue finishes - that's the whole reason you paid for this run.

Install

Manager, search 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

Fully restart ComfyUI and refresh the browser tab. The pack installs nothing - requirements.txt is deliberately inert so an install can't replace ComfyUI's Torch/CUDA stack - which means install problems are almost always a process that didn't actually restart.

Your artifacts land under output/MiniMaxH3/progressive_stage_artifacts. Budget disk accordingly: one completed HIGH latent for a few seconds of H3 video is not a small file, and nothing here deletes anything for you.

Where people get burned

Saving without saving the hash. The path alone is a weaker claim than you think; the pair is what the load node wants. Paste both into a note node in the graph, or as the prefix name - that's what the prefix field is for, and putting the seed and step split in it ("HIGH_8step_low4_tsplit") is the difference between a folder you can read and a folder full of HIGH_001.

Assuming a saved result is a quality endorsement. It isn't, and the author is consistent about this across the whole pack: artifacts record what ran, not whether it was good. If the clip has a seam or the mouth is off, saving it doesn't fix it - and the known limitations are stated openly (dual-pass workflows still have a minor seam colouration, exact voiceprint and word-level timing are not guaranteed).

Forgetting that decoding is a separate stage. If your plan is "save the latent, decode later," make sure your later graph actually contains the decode path - the pack's split philosophy is that generation and decode are separate jobs, and this node follows it.