H3 Chunked v5 · Load Exact Frozen Window (T8 EXP)

The whole reason to freeze a window is that the next run shouldn't need to sample it again. H3 Chunked v5 · Load Exact Frozen Window is that second half: you give it a window you explicitly saved, it verifies the world still matches, and it hands the following window its previous_result.

By manifest path and SHA-256. Not by filename, not by vibes.

What it checks before it gives you anything

Inputs: partial4_denoised_output, lifted_full_av, prepared, plan, expected_window_index (default 0) and the two receipt strings artifact_path and artifact_sha256.

On the way through it re-checks the current source, the global lift, the prepared noise, the plan and the window number against the manifest, and validates the file's bytes against your SHA. Only then does it return cumulative_av_latent, window_result and report_json.

Roughly, a two-window cold graph keeps the verified partial4 source, the global lift and the noise preparation - enough to re-establish provenance - and drops window 0's sampling. That's not just a decode; it's the minimum chain that lets the loader prove nothing drifted.

"Today's model is not claimed equivalent"

That sentence is in the node's own description and it's the most important thing here. The manifest binds an artifact to the execution that produced it. It does not claim that the MODEL and conditioning you have loaded now would reproduce it. This is the opposite of a ComfyUI cache: it's an explicit, auditable handoff between runs, and the pack would rather refuse than flatter you.

Consequences you should plan around:

• Edit the frozen stage's recipe (model, LoRA, prompt) and the correct move is to rerun and save fresh, not to reuse the old file.
• The loader never rewrites or migrates old artifacts. Old saves stay byte-identical; the reader has been taught to understand a couple of precisely reviewed historical manifest formats, and unknown storage versions are still rejected.
• expected_window_index exists so you can't accidentally drop window 0's result into the slot where window 1 belongs.

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 quit and restart ComfyUI, then refresh the tab. This family requires a recent ComfyUI with native H3 support, and the v5 nodes assume the receipt-producing native AV checkpoint Load exists. No extra pip packages for these nodes; the pack keeps requirements.txt empty so an install can never replace your Torch/CUDA. Model weights are the usual set - H3 main model, Qwen text encoder, video and audio VAEs - plus the learned 3D upscaler from the author's HuggingFace.

Where it goes wrong

Blank receipt fields. artifact_path and artifact_sha256 both default to empty. A cold graph built from an example will queue and then refuse until you paste in real values from your own save.

Hand-edited SHA. Copy it exactly. A truncated or retyped hash is a hard failure, and that's the feature.

Wrong window index. Loading window 0's artifact where window 2 is expected is rejected, not reconciled. Pair each load with the right expected_window_index.

Assuming it resumes everything. The load gives you a window; the plan, lift, source verification and noise preparation still run in the cold graph. If your cold graph is missing any of them, you'll get an identity error rather than a shortcut.

Missing nodes. Check for a stale second copy of the pack in custom_nodes shadowing the real install. Rename it with a .disabled suffix - a leading underscore doesn't disable a node pack - restart, and confirm python_module in /object_info.

And the README's own scope note, which is the honest framing for every node in this family: a resume proves the bytes, not the quality. Seams and lip-sync still want your eyes.