MiniMax H3 Smooth Model-Time Bias / 平滑模型时间偏置 (Advanced)

If you've seen the "Detail-Daemon" or "Navyblue" samplers in the image world, you know the trick: nudge what the model perceives as the current noise level so it behaves as if it's at a slightly different point in the schedule - cheaper than changing steps, and often enough to sharpen a result. MiniMaxH3ModelTimeBiasSamplerT8Advanced is that idea, deliberately re-derived for MiniMax H3's shared AV transformer. It biases only the sigma the model sees, inside a smooth tail window, while the integration sigmas and NFE count stay exactly where they were.

Read that once more, because it's the whole mechanism: the actual sampling schedule is unchanged; only the sigma presented to the transformer is nudged. The node outputs a patched model and a ready-made sampler + sigmas, so the cost profile of your run is identical - you're paying the same NFE, just asking the model to read the clock slightly differently. And it adds no random noise; the bias tooltip says it plainly: negative values make the model see a slightly cleaner time.

Inputs

• model and av_latent - your H3 MODEL and joint latent, as usual.
• steps (8), shift_video (12), shift_audio (3) - the sampling contract used to build the schedule.
• bias (−0.05) - the nudge, negative only (range −0.5 to 0). More negative = model sees cleaner time.
• start_progress (0.7) and end_progress (1.0) - the smooth window over which the bias is applied, in progress terms. This is the "tail window": the bias only kicks in during the later, fine-detail phase of sampling, which is where this class of trick does its work.
• bias_domain - video_sigma by default; the domain of the sigma being biased.

Outputs

model, sampler, sigmas - the patched model and its sampler trio, wired into a KSampler-style node. Plus report_json.

Install and honest expectations

Pack install once: ComfyUI Manager → search MiniMax H3 Audio T8, or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into ComfyUI/custom_nodes/, restart. No pip extras, no model downloads.

The author calls this a "scientifically explicit" experiment, and the honest read is: this is a what-if knob, not a proven improvement. The tooltips and README lean hard on "don't expect magic" - one reviewed clip does not establish a general quality win. Start with the default −0.05, keep start_progress around 0.7 so you're only touching the tail, and A/B against the stock sampler on the same seed before trusting it. The neat part is how contained it is: no MODEL wrapper in the "installs a global patch" sense, no step-count change, no noise injection. If it doesn't help, it costs you nothing but a comparison run.