Video quality degradation caused by multi-GPU parallel inference

[wangxihao]

woman_woman_20250514_083244.mp4

Hi, great work, I tried using eight GPUs to parallelize inference with a large resolution, and the inference time did decrease (about 4 times), but the generated video had a fence-like mesh artifact, what could be the reason? As shown in the video and images. Thank you!

[lwmming]

You'd better chunk the audio driven signal correspondingly.

[wangxihao]

You'd better chunk the audio driven signal correspondingly.

In the original fantasytalking code, the text context and audio embedding are cat together, and the text is not cut (in wan21), so I understand the speech without cut.

[lwmming]

The audio signals align with the latents in a frame-by-frame manner, but the text not. so the text has no need to chunk.

[fengye2014]

你好，你是怎么使用8张卡的，需要修改那里才支持多张推理卡，谢谢！

[wangxihao]

你好，你是怎么使用8张卡的，需要修改那里才支持多张推理卡，谢谢！
可以参考wan21.

[SevenMpp]

请问这个问题解决了吗？？？