Zeju1997 / oft

Official implementation of "Controlling Text-to-Image Diffusion by Orthogonal Finetuning".
https://oft.wyliu.com/
MIT License
280 stars 14 forks source link

Where is angular information in the toy experiment? #24

Open Moslem-Sh21 opened 3 months ago

Moslem-Sh21 commented 3 months ago

Hi,

Thank you for your excellent work.

In the toy example in Figure 2 in the paper, it's claimed that the angles (directions) of the neurons are crucial for storing semantic information. However, in the code for "forward_no_norm(self, x)", the following equations are used to process angular information:

x = x / xnorm x = x / wnorm x = x * 60

How is the angular information between x and the weights (filt) represented here? Did I miss something?