MiniMax H3 Audio Refine Dual-Clock Setup / 音频精修双时钟装配 (T8 Advanced EXP)

If the Audio Refine chain is a recipe, this is the bench where the ingredients become exact, measurable quantities. The plan node told you the tail you want; this node takes that signed plan, revalidates it, and assembles every piece a sampler needs to actually run it: the noise, the guider, the sampler, the sigmas, and a latent carrying the audio-only mask. Nothing executes here - it's pure assembly - but by the time it's done, SamplerCustomAdvanced has everything it needs to run the refine tail in one go.

This is also the node where the "dual-clock" concept becomes concrete, and it's worth understanding because it's H3's signature trick. H3 generates video and audio on separate schedules - the video clock and the audio clock run at different effective speeds through denoising, with separate shifts. The pack's validated defaults are video shift 12, audio shift 3; that asymmetry is why a joint model can give you a moving picture and synced sound without one half smearing the other. For audio refine specifically, the mask does the important job: the latent it emits is a nested 0/1 mask where video rows are masked off (0) and audio rows are active (1), so the sampler only touches the sound.

What goes in, what comes out

• plan (H3_T8_AUDIO_REFINE_PLAN), model, positive (CONDITIONING), av_latent - the signed plan and the exact artifacts it was signed against
• Outputs: model, noise (deterministic, seeded), guider (BasicGuider-equivalent), sampler (the stable T8 dual-clock sampler), sigmas (the exact tail), latent (with the nested 0/1 mask), and report_json

Wire those five into SamplerCustomAdvanced in the standard positions, then decode with the H3 audio VAE and feed the result to the Quality Gate.

Installing it

Part of the T8mars/comfyui-minimax-h3-audio-T8 pack - search "MiniMax H3 Audio T8" in ComfyUI Manager, install, restart, or:

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

Update ComfyUI itself first (recent core: comfy_api.latest, comfy.weight_adapter, comfy.patcher_extension, comfy.ldm.minimax); requirements.txt is intentionally empty. And because this is an EXP chain, start from the bundled workflow - hand-wiring SamplerCustomAdvanced from scratch is where people lose an afternoon.

Where people get burned

The silent killer is treating this like a generic setup node and feeding it a plan that wasn't signed by your actual audit. The revalidation is there for a reason: change the model, the conditioning, or the latent between the audit and this node, and the assembly should refuse rather than run a tail on artifacts it wasn't designed for. Don't bypass it with "it'll probably work." Also remember the pack's hard rule against stacking multiple sampler/attention/MODEL-forward owner nodes - this node emits a specific dual-clock sampler; if you override it with another sampler downstream you're back to fighting the architecture. If a run comes out with mangled audio, your first suspects are a denoise value you pushed too high in the plan, or a mask/latent that doesn't match the video stream's timeline.