Closed heart-du closed 4 months ago
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。
I'm not sure if the ADD paper mentions whether the teacher model uses CFG. Could you please tell me the specific location?
我看别人复现的ADD代码里面有
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
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