MiniMax H3 Modality-Stable AV Noise / AV独立随机噪声 (Advanced)

In diffusion, the seed determines the noise, and the noise determines everything. That's a problem for A/B testing: if you change the video canvas to compare resolutions, and the audio noise stream silently shifts too, then a "resolution difference" is really a "different noise all over" difference, and your comparison is junk. MiniMaxH3SPEEDModalityStableNoiseT8Advanced is a tiny, one-input node that fixes that specific failure: it generates deterministic video and audio noise from separate derived seeds, so changing only the video canvas can't silently change the audio noise stream. It's the control-subject node of the SPEED diagnostics kit.

How it works

You give it one thing - seed (default 2608184001, with a control_after_generate widget) - and it returns a single noise object. Internally the pack derives the two modality streams from that one seed (there's a hard-coded constant in the source, SPEED_AUDIO_NOISE_SEED_XOR, that perturbs the audio branch), so the video noise and audio noise are deterministic and independent of each other. Resize the video canvas for a SPEED stage and the video noise reshapes as it must - but the audio stream, having no spatial dimension to stretch, stays bit-for-bit the same. That's what "modality-stable" means here.

If that sounds niche, that's because it is. The description says it straight: "Intended for controlled H3 SPEED baselines and diagnostics." You reach for this when you're building a comparison - SPEED vs. straight sampling, manual-sigma A/Bs, spectrum experiments - and you want the audio held constant so the only variable is what you're actually studying.

Wiring it up

The noise output plugs into the sampler that accepts a noise input. In the SPEED graph, that's the SPEED sampler's noise slot (or a stock H3 sampler when you're doing the baseline side of the comparison). Keep the same seed across both sides of an A/B and the video noise matches too - this node is at its best when you use it consistently, not as a one-off. A couple of practical notes:

• The default seed being a fixed magic number is deliberate - it's the pack's standing reference point for reproducible tests. Change it when you want fresh content, keep it when you want to compare runs.
• Because it's pure math (no model, no GPU load beyond building the noise tensors), it's effectively free to run. There's no reason not to use it in a controlled experiment.

Installing

Standard pack install: ComfyUI Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes, restart. No pip deps, no model files, no API. This is one of those nodes that looks like a toy until you've burned an afternoon chasing a "regression" that was actually just seed drift - after that, modality-stable noise feels like a necessity. It's experimental, but it's also about the least risky experimental node in the pack: the worst it can do is give you noise you didn't want, which a seed change fixes instantly.