H3 Chunked v1 · Load Exact Frozen Segment (T8 EXP)

The pitch is simple: you already rendered segments 0 and 1, you changed something about segment 2, and you don't want to pay for the first two again. H3 Chunked v1 · Load Exact Frozen Segment is the other half of the freeze/resume pair - it reads a segment you explicitly saved, re-verifies that the world still matches, and hands the next PASS2 segment its previous_result without resampling anything.

Explicitly. By SHA. If that phrase sounds repetitive, it's because the pack says it in every description, and it means it.

How the load works

You supply artifact_path and artifact_sha256 - copied verbatim from the save node, not reconstructed by hand - along with the current source_segment, segment_spec, pass2_context and plan. The node then does two things at once:

1. Reads the manifest and checks the file's bytes against the SHA you gave it.
2. Re-checks the current graph's source latent, plan and segment identity against what the manifest recorded.

Only if both agree does it return cumulative_av_latent and segment_result, ready to be fed into the next segment's previous_result. It also rebinds the current original full audio, which matters on this route because v1 PASS2 keeps first-pass audio as the final track.

The pay-off: for a three-segment chain, the cold-resume graph can drop the whole earlier part - the first pass and the earlier segments' samplers - and just decode and run the remaining windows. The pack's own example set includes a cold-resume variant per saved state for exactly this reason, and its CPU-side test suite covers "complete save, then a fresh executor loads segment N and runs only the tail".

What it is not

It is not an automatic cache. The manifest binds bytes, not behaviour. If you changed the MODEL, the conditioning, the LoRA stack or the prompt after freezing, the load does not treat that as a cache hit, and it isn't claiming today's recipe would have produced the frozen result. It's the pack's standing rule: the file proves what you saved, not what you'd get now. Change the frozen part of the recipe and the honest move is to rerun and save again.

It doesn't delete or rewrite anything. Old artifacts stay byte-identical; the node is read-only, and the storage layer will not silently migrate your old files forward.

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 after install or update, then refresh the browser tab - "restart the queue" is not a restart. This family needs a recent ComfyUI with native H3 support. Base and EXP nodes install with no extra pip packages; the pack's requirements.txt is empty on purpose so it can never overwrite ComfyUI's Torch/CUDA. Main model, Qwen text encoder and the two VAEs are separate downloads, listed in the README's model table.

Where it goes wrong

Empty path or SHA. Both fields default to empty strings. A cold-resume graph will queue and then refuse; fill in real receipts before hitting run.

Retyped SHA. Copy-paste it. A truncated or hand-edited hash fails the check - which is the correct behaviour, and better than discovering at frame 400 that you loaded the wrong segment.

Wrong pairing. The load wants the save from that segment, plus the current segment's spec and plan. Loading segment 1's artifact into the slot expecting segment 0's is rejected, not reconciled.

Stale duplicate pack. If the node is missing after updating, check for a second copy of the pack in custom_nodes - old test copies serve old schemas. Rename with a .disabled suffix (leading underscores don't disable), restart, and verify the python_module in /object_info.

One more, from the README's own caveat list: passing one sample doesn't certify your asset, your GPU or a longer window. The frozen segment will resume exactly; whether the seam looks right is still your eyes' problem.