search

nupurkmr9 / concept-ablation

Ablating Concepts in Text-to-Image Diffusion Models (ICCV 2023)
https://www.cs.cmu.edu/~concept-ablation
MIT License
147 stars 20 forks source link

FID scores #4

Closed jijunhao closed 1 year ago

jijunhao commented 1 year ago

This is a good jod！Great work! I am inspired by it. However, I have some minor issues. I'm curious about the difference between stable diffusion of CompVis and the diffusers in this repository. Are they different implementations of the same method? Another concern is whether fine-tuning ，even with regularization，would affect FID scores for other categories？

Thanks.

nupurkmr9 commented 1 year ago

Hi, thanks for the interest in our work and sorry for the delayed response.

The CompVis and diffusers are different implementations of the same method. The diffusers version also consists of chatGPT API to generate the prompts corresponding to the concept being ablated whereas CompVis implementation requires a text file with those prompts as an argument. Thus it might be easier to use one or the other implementation for ablating a new concept based on whether you want to provide your own prompts or have a chatGPT access key.

Regarding the FID scores for other categories, we calculated the FID on MSCOCO dataset and found no significant difference. Thus the ablation process probably doesn't affect the unrelated concepts.