CLIP Guided Diffusion
Example report comparing two runs, one with `timestep_respacing` at 100, the other at ddim100. The caption is "a red ball". It seems DDIM is better in this instance, although in general going beneath 250 timesteps is risky. You can use the slider to examine the difference between the current noisy sample and the current denoised prediction. CLIP sees a variety of cutouts of the final blending of the two images, blending increasingly more of the noisy and increasingly less of the denoised over time.
Created on October 1|Last edited on October 3
Comment
Run set
2
Run set
2
Add a comment