Dual-clock sampling: separate video and audio step counts — MiniMax H3 tutorial
Control video and audio sampling separately in ComfyUI and compare it carefully with an acceleration LoRA.
Audience: Advanced users balancing sound quality, sync, and generation speed.
Hardware: An NVIDIA environment capable of running H3 in ComfyUI.
Prerequisites
- The T8 audio-video nodes
- A native H3 workflow
- A fixed audio-video test prompt
Steps
- Confirm the official H3 ComfyUI baseline outputs an MP4 with audio, then save its seed, prompt, canvas, frame count, and 20-step result.
- Find MiniMax H3 Audio T8 in ComfyUI Manager or install it with the comfy CLI, then fully restart ComfyUI.
- Open examples/workflows/01-basic-generation and replace the models and media. Start with 22 frames, a small canvas, and the stock sampler.
- Wire MiniMaxH3AudioConditioningT8, MiniMaxH3DualClockSamplerT8, MiniMaxH3AVDecodeT8, and MiniMaxH3PreflightT8 first. Run preflight and fix size, tag, or VRAM errors.
- Hold video steps constant while changing the audio clock, then reverse the test. Record total time, speech clarity, event sync, and output stream metadata each time.
- Add Turbo LoRA only after the stable path passes. Test experimental nodes, SageAttention, and long-video graphs separately rather than stacking them in the first run.
Commands
comfy node install minimax-h3-audio-t8
cd ComfyUI/custom_nodes && git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8
Caveats
- Re-test community speed claims on your own hardware and keep experimental nodes separate from stable ones.
View original source · 0xMomo · 2026-08-23
Guide · Source checked: 2026-08-23 · Site testing: No generation test performed
Check the prerequisites, then follow the author for the full demonstration and updates.
Follow the original source for its materials. Missing assets or settings are not reconstructed here.
Applicable versions
MiniMax H3 Audio T8 1.44.0 · ComfyUI native H3
Troubleshooting
prompt media tag validation failed
Match <Picture 1>, <Video 1>, and <Audio 1> numbering to the wired inputs; do not omit indices with multiple references.
OOM, NaN, or missing audio
Save the complete Error Report, retry at 22 frames on a small canvas with stable nodes, and remove experimental patches and stale caches.