H3→LTX Learned · Audit Completed Refiner (T8 EXP)

How it differs from the plain audit

There are two audits on the H3→LTX learned route and they are not redundant.

The plain one (MiniMaxH3LTXLearnedStageAuditEXPT8) rechecks the source, the adapter report and the connected sampler controls, then admits in its own report that it does not prove the external sampler executed. Its report literally carries sampler_execution_proven: false.

This one closes that hole. Same input list, plus one extra: sample_proof, the live-only evidence handed out by MiniMaxH3LTXLearnedStageSampleEXPT8. It rechecks the "old source-bound candidate contract" and requires the matching live Stage Sample proof. If the proof came from a different run, a different process, or doesn't correspond to the candidate, it raises.

So: plain audit = "the wiring is the wiring I signed." Certified audit = "and the LTX sampler actually ran for this candidate."

When to reach for it

Reach for the certified version when you intend to act on the result - freeze it, chain it into a downstream stage, hand it to someone else, or add a stage effect on top and want the effect's coverage to be observable. The pack's own LTX campaign notes show what happens without that discipline: graph edits that look cosmetically identical, sampler modes changed between an A and a B run, and results that differ numerically with no record of why.

Skip it when you're iterating and just want to see pixels. It's more inputs to rewire every time you change something, and every one of those inputs is a claim that has to stay true through the run.

What it does not give you, and the node's own description is explicit: no portable cache, no quality claim. A passed audit is not "this render is good." It's "this candidate provably came out of the stage you configured."

Inputs and outputs

Inputs are the full audit set plus the proof: stage_boundary, h3_latent, original_h3_av, ltx_video_latent, adapter_report_json, model, noise, guider, sampler, sigmas, setup_report_json, candidate_latent, sample_proof. Nothing to tune.

Outputs: candidate_latent (onward to LTX decode), original_h3_av (onward to the H3 audio VAE decode - the audio never entered LTX), duration_seconds, report_json.

Bind → Sample (produces sample_proof) → Certified Audit → TAEHV decode
                                       ↘ original_h3_av → Original Audio Decode

Install

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

Or ComfyUI Manager → search MiniMax H3 Audio T8. Fully exit and restart ComfyUI afterwards, then refresh the page. If the node stays missing while GitHub has it, your Manager hasn't picked up the release yet - clone manually, since registry publication and GitHub releases are independent.

Model side: H3 transformer in models/diffusion_models, Qwen3-VL text encoder in models/text_encoders, H3 video and audio VAEs in models/vae, plus separate LTX-2.x weights, its text encoder and the learned adapter. No package installs come from this repo - requirements.txt exists explicitly so nothing can overwrite ComfyUI's torch/CUDA build.

One licence note that applies to this entire family: the local H3 weights are only licensed outside the US, EU, UK and Republic of Korea, and the exclusion covers outputs, not just weights.

Common issues

"Audit received another stage recipe" or a proof mismatch. You're validating a candidate with a proof from a different sampling run. The proof is live-only by design and doesn't survive a restart - rerun both nodes in one pass.

Works in the warm graph, fails after a restart. Expected. Cold/warm parity on this route depends on the boundary inputs being rebuilt identically, and the docs record that cross-process numeric differences are a real thing on this pipeline rather than a bug to paper over.

Red nodes. Update ComfyUI core, frontend and Manager together, then restart. Half-updated installs are the usual cause on this pack.