MiniMax H3 audio quality regression after upgrading to ComfyUI v0.31.0
#4
by starsfriday - opened
After upgrading ComfyUI from v0.30.0 to v0.31.0, MiniMax H3 still generates videos successfully, but the audio becomes noisy and loses stereo coherence.
The same workflow, model, prompt, seed, and sampling settings were used for all tests.
Before upgrade — normal audio
https://github.com/user-attachments/assets/3ddfd8e2-32e9-470b-95af-52e88db277d9
After upgrade — broken audio
https://github.com/user-attachments/assets/e1d9b8bb-5a08-4cb0-a05d-3972e90d1117
After restoring legacy audio sampling — fixed audio
https://github.com/user-attachments/assets/3c2d3fe9-09aa-4bc5-bd80-757d7c4345e7
The issue appears to be related to the MiniMax H3 sampling changes introduced in:
https://github.com/Comfy-Org/ComfyUI/pull/15243
A temporary custom-node workaround is available here:
https://github.com/starsFriday/ComfyUI-MiniMax-H3-LegacySampling