推理的时候为什么没用CFG呢？

Question

Closed this issue 5 months ago · 5 comments

推理的时候为什么没用CFG呢？

Answer 1 · 2024-04-12T09:13:44.000Z

We find that using cfg during inference may introduce the artifacts, which also aligns the settings in SDXL-Turbo, they set the guidance_scale=0

Answer 2 · 2024-04-12T12:59:55.000Z

We find that using cfg during inference may introduce the artifacts, which also aligns the settings in SDXL-Turbo, they set the guidance_scale=0

可是ADD在训练的时候是把CFG蒸馏到学术模型中去了，AddSR似乎并没有，但是效果却也很好，挺奇怪。

Answer 3 · 2024-04-12T14:18:24.000Z

I'm not sure if the ADD paper mentions whether the teacher model uses CFG. Could you please tell me the specific location?

Answer 4 · 2024-05-14T02:08:53.000Z

我看别人复现的ADD代码里面有

Answer 5 · 2024-07-15T08:27:18.000Z

我看别人复现的ADD代码里面有

请问有别人复现的ADD代码吗，可否发个链接，谢谢