H3→LTX Learned · Sample Separate Refiner (T8 EXP)

What it is

Functionally this is SamplerCustomAdvanced: you connect a NOISE, a GUIDER, a SAMPLER, SIGMAS and a LATENT, and it runs one native Core sampler call. It returns output (the finished latent), denoised_output (the model's predicted clean latents), a sample_proof, and report_json.

The difference is the sample_proof. This node records live-only evidence of the step callbacks it observed - the pack describes it as proof of three step callbacks, which lines up with the distilled LTX stage-2 graphs that run three steps - and hands that evidence forward. Feed it to MiniMaxH3LTXLearnedStageCertifiedAuditEXPT8 and the audit will require the proof to match its own contract.

"Live-only" is the important word. That proof is a receipt from this process and this run. It is not a cache token you can save, move between machines, or revive after a restart. If you want the other direction - freezing the input data itself so a later run doesn't repeat the H3 half - that's the RGB source save/load pair and its own storage format, not this.

Where it fits

In the learned H3→LTX route the order is roughly:

Bind (signs adapter output + sampler controls)
  → MiniMaxH3LTXLearnedStageSampleEXPT8   ← you are here
  → MiniMaxH3LTXLearnedStageCertifiedAuditEXPT8
  → LTX video decode (TAEHV etc.)

You can absolutely use a plain SamplerCustomAdvanced instead. You'll get the same sampling and lose the proof, so the Certified Audit can't be used downstream and the plain Audit will still work fine. That's the trade: use the pack's sampler when you want the graph to be able to prove it ran the LTX pass.

Inputs and outputs

• noise - an LTX-appropriate NOISE (RandomNoise or the pack's noise wrapper, seeded and with the batch index you intend).
• guider - CFGGuider or BasicGuider for the LTX stage. LTX distill graphs are typically CFG 1.
• sampler - a Core sampler object.
• sigmas - the LTX schedule. Shape and dtype are part of what gets re-checked upstream, so don't rebuild it casually between warm and cold runs.
• latent_image - the ltx_video_latent that came out of the Bind node. Not an H3 latent, not an AV latent.
• stage_boundary - the signed contract from Bind.

Outputs: output is the one you decode. denoised_output is the x0 prediction - the pack is consistent across its families in warning people not to feed denoised_output into a handoff that expects a terminal result, and this is one of those. sample_proof goes to the Certified Audit. report_json is for your own log.

Install

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

ComfyUI Manager: search MiniMax H3 Audio T8, install, then quit ComfyUI completely and restart, then refresh the browser. Registry and GitHub release on separate schedules; if Manager looks stale, clone from GitHub.

Beyond the H3 set (transformer in models/diffusion_models, Qwen3-VL encoder in models/text_encoders, video/audio VAEs in models/vae), this route needs the LTX-2.x weights, its encoder and the learned adapter, each with its own licence and download. The pack itself ships no weights and no Python dependencies - requirements.txt is intentionally empty of packages.

Start from a bundled graph under examples/workflows/35-h3-ltx-latent; these EXP routes are documented as "use the matching workflow", and the wiring around this node has more moving parts than the node itself.

Common issues

Sampler runs but the Certified Audit rejects. The proof is per-run. If you swapped in a cached latent, ran the graph in a way that skipped the sampler, or restarted the process and tried to reuse an old proof, the audit is supposed to refuse. That's the certification doing its job, not a bug.

LTX output looks soft or plasticky. Before blaming the node: LTX-2.x quality is sensitive to sigma schedule and step count, and the local community's own verdict on LTX has always been "fast, improving, not the fidelity king." Three distilled steps is a preview-grade refinement, not a restore.

Red nodes / missing node. Update ComfyUI core, the frontend and Manager together and restart fully - a partial update is the usual culprit on this pack. If it's still absent, your core predates the native H3 support these nodes bind to.