H3 Chunked v5 · Verify Native Partial4 Source (T8 EXP)

The standard joint 4+4 recipe for MiniMax H3 in this pack works like this: run the first four steps of an eight-step schedule at low resolution, upscale the whole AV latent once with the learned 3D upscaler, then run the remaining four steps per time window. Two-pass high/low noise sampling is familiar enough from Wan 2.2 - same idea, different plumbing, because H3's audio rides in the same latent.

The failure mode is obvious once you say it out loud: if that low-resolution "partial4" latent isn't actually the denoised output of a real, completed native checkpoint, every window you sample on top of it is decorating garbage.

H3 Chunked v5 · Verify Native Partial4 Source is the gate. It takes the receipt outputs of the native AV checkpoint Load - av_latent, resume_verified, checkpoint_id, content_sha256, file_sha256, manifest_json and report_json - re-checks the external manifest, and then strips only the Load provenance so the original partial4 source identity can bind the v5 windows.

That last part is the subtle bit. The v5 chain wants a source that looks like a completed sampling state, not one with a checkpoint loader's fingerprints all over it. This node is the adapter that produces the honest version of that: verified, then re-presented as a source.

What comes out

partial4_denoised_output - the LATENT that feeds the rest of the v5 chain, starting with H3 Chunked v5 · Global Learned Lift. And verification_report_json, which you should actually read: it's where the manifest comparison and the identity rebinding are written down.

Note there's a second, unrelated "load a partial4" node in the family, MiniMaxH3ChunkedV5WindowLoadEXPT8 - that one resumes completed windows, not this initial source. Different job, similar name, easy to mix up.

Wiring the rest of the chain

Rough order, all on the v5 side:

1. Native AV checkpoint Load → this node → partial4_denoised_output.
2. Global Learned Lift → lifted_full_av + lift_receipt.
3. Prepare Full AV Noise → prepared (one global video+audio noise draw, shared by every window).
4. One Joint AV PASS2 Window per window, with previous_result chaining them.
5. Optional Window EAV / Project External Relay per window, each with its own audit.
6. Freeze ONE Joint AV Window / Load Exact Frozen Window if you want to stop paying for work you already did.

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 browser tab. This pack needs a recent ComfyUI with native H3 support - the v5 nodes assume the native AV checkpoint loader exists, since that's where all these receipt fields come from. No extra pip packages: the pack keeps requirements.txt deliberately empty so an install can't replace ComfyUI's Torch/CUDA stack. Weights (main model, Qwen, video VAE, audio VAE) are separate; the learned 3D upscaler used by the next node comes from the author's t8star HuggingFace repos.

Where it goes wrong

You wired a plain LATENT instead of the receipts. Every receipt output of the Load has to be connected - the node is checking a manifest, not just a tensor. A tensor alone gives it nothing to verify against.

Resume flag false. resume_verified is an input here, and the node means it. If your save didn't verify on load, this is not the place to argue about it.

Confusing v1 and v5. The older chunked route slices a first-pass latent per segment and lifts per segment. The v5 contract lifts globally once and rejects the per-segment source path - do not mix nodes from the two families and expect a coherent result.

Stale duplicate installs. If inputs don't match the example graphs, look for a second copy of this pack under custom_nodes. Old test-copy folders serve old schemas; rename them with a .disabled suffix (a leading underscore doesn't disable a pack), restart, and check python_module in /object_info.

Finally, the README's standing warning applies with extra force here: EXP workflows are validated on specific samples. Verification proves your bytes are the bytes. It doesn't prove your clip looks good.