MiniMax H3 Scheduled Drive Audio Injection (T8 Advanced)

H3 can take an audio track as a drive - a conditioning signal that steers the joint generation - but how hard that drive bites depends on when during denoising it's applied. This is an experimental dual-clock sampler that repeatedly re-anchors your complete drive-audio latent onto its own sigma path, across a window you control. Instead of the drive being a one-time setup baked in at the start, it keeps getting re-presented to the model as the noise schedule progresses, which is a stronger and different lever than stock behavior.

Read the name carefully: drive audio, not source audio, and not a mix. It doesn't isolate speech from music or effects - the whole supplied audio latent gets anchored, warts and all. If you want your original dialogue preserved bit-for-bit, the README points elsewhere (lock_source/mux_audio routes). This node is for when you want the audio to influence generation more aggressively.

How it works and the inputs that matter

It's a sampler factory: feed it model, the nested av_latent, the drive_audio (AUDIO), and an audio_vae to encode it, and it returns model, sampler, sigmas for a KSampler plus the processed av_latent and a mux_audio output.

The controls that actually matter:

• mode - report_only (default) is an unchanged dual-clock bypass. Injection only happens when you switch out of it. This is the pack's standard "opt in or the node is a no-op" pattern.
• start_percent / end_percent - the injection window across the denoise. 0 is the first/high-noise step, 1 is the last/low-noise step. Default full range 0→1; shrink it to inject only early (structure) or only late (detail).
• strength (1) and envelope (constant) - how hard, and whether that strength ramps.
• injection_seed - independent seed for the re-noise at each anchor.
• lock_final_audio - if true, replaces the final generated audio latent with the encoded drive audio. The tooltip is blunt: for unchanged source PCM, use mux_audio, not this.
• allow_unverified_patch_stack - false by default; it blocks combinations of patches that haven't passed the validation matrix. Leave it off.
• optional scheduler (native_flow default) and final_audio.

Installing and the honest take

Part of T8mars MiniMax H3 Audio T8:

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

or ComfyUI Manager → "MiniMax H3 Audio T8", restart. No pip deps; it runs on your existing model/VAE.

Honestly, treat this as a power-user experiment. The default is a clean bypass, so you can drop it in, confirm your existing workflow is untouched, then start tightening the window. Where it's interesting is on a clip whose audio feel keeps drifting from the reference - re-anchoring mid-denoise is a real difference from one-shot conditioning. Where it will bite: lock_final_audio sounds like the easy win but can fight the video side, and stacking it with other experimental patches is what allow_unverified_patch_stack is there to stop. Start with report_only, then a window like 0→0.6, and judge by ear.