About the jointattn layer

[xxlong0]

About the jointattn layer

[flamehaze1115]

Hi Xiaoxiao,

thanks for your great work!

I have a question about the jointattn layer for normal and rgb images. I found that you didn't use any jointattn layer in the released model, which is different from the paper. Just curious why this still works without the jointattn layer and will the performance becomes better without it?

Hello, thanks for your interetes in our work. Due to some constraints of company, we cannot provide the complete model weights and training codes at this time.
Although the joint cross-domain attention is not activated, the incomplete model still works but is less better than the complete model with cross-domain attention in some cases. Due to our domain switcher design and joint training, the two domains can be roughly aligned in the learned latent space, which should be the potentials of diffusion models. With a cross-domain attention, the consistency of the two domains can be further improved.

[yf1019]

Hi

Hi Xiaoxiao,
thanks for your great work!
I have a question about the jointattn layer for normal and rgb images. I found that you didn't use any jointattn layer in the released model, which is different from the paper. Just curious why this still works without the jointattn layer and will the performance becomes better without it?

Hello, thanks for your interetes in our work. Due to some constraints of company, we cannot provide the complete model weights and training codes at this time. Although the joint cross-domain attention is not activated, the incomplete model still works but is less better than the complete model with cross-domain attention in some cases. Due to our domain switcher design and joint training, the two domains can be roughly aligned in the learned latent space, which should be the potentials of diffusion models. With a cross-domain attention, the consistency of the two domains can be further improved.

Hi Xiaoxiao,
thanks for your great work!
I have a question about the jointattn layer for normal and rgb images. I found that you didn't use any jointattn layer in the released model, which is different from the paper. Just curious why this still works without the jointattn layer and will the performance becomes better without it?

Hello, thanks for your interetes in our work. Due to some constraints of company, we cannot provide the complete model weights and training codes at this time. Although the joint cross-domain attention is not activated, the incomplete model still works but is less better than the complete model with cross-domain attention in some cases. Due to our domain switcher design and joint training, the two domains can be roughly aligned in the learned latent space, which should be the potentials of diffusion models. With a cross-domain attention, the consistency of the two domains can be further improved.

Got it. Thanks so much for your prompt response~