search

ToTheBeginning / PuLID

Official code for PuLID: Pure and Lightning ID Customization via Contrastive Alignment
Apache License 2.0
1.23k stars 74 forks source link

question about fig2 in the paper. #27

Open dingangui opened 4 months ago

dingangui commented 4 months ago

Hello, thanks for your incredible work!

In the 'Accurate ID Loss' section in the bottom right corner of Figure 2 of the paper, there are two generated images both denoted as 'predict x_0'. Are both of these images produced by the Lighting T2I? I guess they represent T2I w/ ID and T2I w/o ID, respectively. However, upon closer inspection, it appears that the IDs of both images are well-preserved, which contradicts my speculation. What are these two images' actual meanings and why do you connect them with a vertical line?

image
guozinan126 commented 4 months ago

When calculating the ID loss, the two images involved in the calculation are both generated by the Lightning T2I training branch, and both images are generated under the T2I w/ ID setting. Additionally, only when calculating the Alignment loss, the contrastive pair is composed of T2I w/ ID and T2I w/o ID.

Luh1124 commented 2 weeks ago

When calculating the ID loss, the two images involved in the calculation are both generated by the Lightning T2I training branch, and both images are generated under the T2I w/ ID setting. Additionally, only when calculating the Alignment loss, the contrastive pair is composed of T2I w/ ID and T2I w/o ID.

Is there no id loss calculated with the given identity images here? Is it only done on the generated images? Also, I would like to ask, how is it ensured that the ID loss done by ArcFace is differentiable for gradient backpropagation?