great work!
#1
by pr1vatejoker - opened
Great work! The visual quality is really good, especially the realism and facial quality. The only issue I'm having is with the audio. I'm using minimax_h3_fl2va_pruned_int8_convrot.safetensors, which I assume is the model most people are using.
There's no reason to manually type in the sigmas....that's very close to the "simple" scheduler. Anyway, I've tried it in euler and res_multistep and it still very much looks like the full quality model at https://huggingface.co/Kijai/MiniMax-H3_comfy/tree/main/loras using pruned int8 convrot is king.
By the way, do not merge the LoRAs in any case into said convrots as that will destroy quality.