Meridian 4 · 原生3前向生成（T8）

What it is, and why it's the last node in the chain

This is where the Meridian chain actually spends money. Node 1 configured the resources, node 2 built the material, node 3 wrote the camera and time tracks - this node takes that prepared plan and produces the MP4. It's marked as an output node, so it's what ComfyUI executes when you hit Run.

It is not a sampler-family node you can drop into a normal H3 graph. It loads the converted Meridian checkpoint itself, runs the native H3 joint AV pathway with Meridian's merged-DMD recipe, decodes with the native H3 video VAE, and delivers a file. You keep the camera plan; the sampling decisions are already made for you.

How it works

The checkpoint here is a 34 GB native ConvRot INT8 conversion with DMD merged in. That matters twice: DMD is why there are three forwards instead of a twenty-step schedule, and ConvRot INT8 is why this can run on 20- and 30-series cards that have no fp8 acceleration at all. Before it samples, the node checks that the loaded model really has its 250 INT8 ConvRot layers - if a different file got substituted, it raises instead of quietly producing garbage.

The recipe is fixed: CFG 1, Euler, video and audio shift 3/3, three forwards, joint audio+video in the native layout. DMD is a distillation method, and the general rule for distilled checkpoints is CFG 1 with a tiny step count; the sharp edge is that one sampler setup means no rescue if a specific shot lands badly.

Delivery is deliberately paranoid. It decodes through the native VAE, checks the frame count and dimensions, writes an H.264 MP4 to a partial file, then fully re-decodes it - verifying codec, dimensions, frame count and 24 fps PTS spacing - before publishing a report and the file under your output directory. If verification fails you get no clip and a job report explaining which stage broke, which is a lot better than a half-broken rollout.

The inputs that matter

• prepared - the plan output from the camera node. No plan, no generation.
• seed - unsigned 64-bit, default 1234. Same seed, same plan, same assets: repeatable.
• audio_mode - silent or source_1to1. Freeze or speed-changed camera work belongs in silent. source_1to1 muxes the original audio window back in, and the author is explicit that this delivers the source track rather than certifying newly generated lip sync. It also requires VIDEO material with a genuine 1:1 source time mapping - pick it with an IMAGE input and you get an error.
• output_directory - blank writes to output/meridian/. Otherwise put it somewhere with room for the file and its JSON receipt.

Outputs: video (VIDEO, into Save Video, upscaling or interpolation), saved_path (STRING) and report_json (STRING). The report is written as the job progresses, so after a crash it still tells you the last stage reached.

Installing

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

Restart fully and refresh. Manager finds the pack as MiniMax H3 Audio T8. You need a recent ComfyUI with native H3 support and native ConvRot INT8 loading - an older build will fail to load this checkpoint. Weights come from t8star/Meridian-Comfy; merge its models/ tree into ComfyUI/models/, and don't add the DMD LoRA, because it's already inside the file. The pack ships no pip dependencies of its own on purpose, so install shouldn't disturb your Torch stack.

Where people get burned

A silent delivery contains no audio stream at all - that's enforced, not a bug. If you expected sound from a silent run, you asked for the wrong mode.

Files come out as Meridian_<job>.mp4 plus a JSON. The partial file stays a partial until verification passes; don't go hunting for footage in .partial files.

Cancellation follows Core's normal interrupt, not a private panel. It doesn't touch your queue or other jobs.

Long windows and high resolutions still OOM. INT8 makes the model fit; it doesn't make arbitrary length free. And per the pack's own README, tested samples passing is not a promise for your footage, your card or your window size - the workflow ships with specific tested recipes for a reason.