BrushNet icon indicating copy to clipboard operation
BrushNet copied to clipboard

So there is no attention block in brushnet right? no self-attn and cross-attn

Open yunniw2001 opened this issue 6 months ago • 0 comments

In the paper, it says "To process the masked image features, BrushNet utilizes a clone of the pre-trained diffusion model while excluding its cross-attention layers. The pretrained weights of the diffusion model serve as a strong prior for extracting the masked image features, while the removal of the cross-attention layers ensures that only pure image information is considered within this additional branch." So I assume that brushnet only keep self-attention block.

But when I check brushnet config file and code, I print brushnet modules, and only see some resnet block, linear, etc. And in brushnet config file, the 2D block it specified is pure conv block(DownBlock2d, Mid, Up).

So the cross attention you remove is not only cross attention layer, but also self attention layer, which is 'CrossAttnDownBlock2D''s crossattn block, right?

yunniw2001 avatar Aug 07 '24 04:08 yunniw2001