openai / guided-diffusion

MIT License
6.03k stars 803 forks source link

Non-Square Images #83

Closed Squeemos closed 1 year ago

Squeemos commented 1 year ago

I'm using this unet architecture to create images, but I'm curious about using non-square images. Current the only problems seems to be in the output_blocks, specifically with th.cat([h, hs.pop(), dim=1]) With non-square images the dimensions aren't correct, which is understandable.

Is there a good way to go about getting this setup to use non-square images? What's interesting is inside the u-net init it takes in an image size but doesn't use it anywhere else

tevkhieu commented 1 year ago

Did you find a way to create non-squared images?