刚测试 int8剪枝版有伪像和杂音，这个底模式必须bf16吗
如题，谢谢
It is based on the BF16 framework for distillation. Therefore, for the tests, we also sampled the BF16 model for inference.
It is based on the BF16 framework for distillation. Therefore, for the tests, we also sampled the BF16 model for inference.
了解，感谢！！
It is based on the BF16 framework for distillation. Therefore, for the tests, we also sampled the BF16 model for inference.
I tried your example workflow, but I still get distorted sound, just like in the LightX2V LoRA example video of your comparison group.
I used minimax_h3_fl2va_pruned_bf16.safetensors, and the only difference was that I changed qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors to qwen3vl_32b_minimax_h3_int8_convrot.safetensors.
I found that the video quality and details are much better than LightX2V, so it would be great if I could get the audio working properly.