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

Training and evaluation process? #20

Closed DaShenZi721 closed 7 months ago

DaShenZi721 commented 7 months ago

Hi, thanks so much for offerring this repo!

During the training process, do I need to individually train model on each of the 750 prompts?

During the evaluation process, for each prompt, I calculate the best score among the images generated at different training steps, and considered this best score as the best score for that prompt. After obtaining the best scores for all 750 prompts, I calculated their average as the final result. Is this the calculation process for Table 1 in the paper?

Zeju1997 commented 7 months ago

We did train on each of the 750 prompts to get the score as in the paper, but you could train on each subject and just sample the 25 text prompts. Yes, it should be consistent with the Table 1.