search

subrtadel / DIA

MIT License
15 stars 3 forks source link

Token regeneration results #5

Open edward3862 opened 11 months ago

edward3862 commented 11 months ago

Hello, I'm running the provided script for CLIP feature estimation. When I regenerate the images by setting regenerate_tokens, I get the following results in token_regeneration_with_ref.png.

image image

I'm not sure whether this meant to be correct. In my understanding, the regenerated images are supposed to contain similar content with the input image. Is there anything I missed and how should I fix this issue?

Thanks!

subrtadel commented 11 months ago

You are right, the images should contain similar content. Unfortunately, the CLIP feature estimation is very sensitive to hyperparameter setting. I would try increasing the number of optimization steps (conditioning_optimization.opt_iters) in parameter_estimation.yaml.

dsubias commented 7 months ago

Hello,

What is the number of "opt_iters" used to obtain the results shown in the teaser of the article?

Thanks!