Code and Script

[ms-ko]

Thank you for sharing your work!

1. The LearningRateMonitor is not defined in train.py, causing a Callback error [Compvis].
2. "Bilnear" in finetune_data might need to be changed [Compvis].
3. The Laion-retrieval currently seems to be not working [Diffuser].
4. For nudity removal (I followed the setting in Diffuser), I generated 200 prompts with people/body based on your llm generation code and used them for class-prompt, setting the caption_target as "nudity, nsfw" as suggested by the authors. This will generate 1,000 safe images in the data path. Does it work correctly? or did I miss something?

[nupurkmr9]

Hi, Thanks for the interest in our work and your queries.

1. We are not using LearningRateMonitor but will also remove this from the code.
2. You are correct, Lanczos will be better than bilinear. Will update to that in the next commit.
3. I am aware of the issues with laion-retrieval not working. In the meantime, we have also provided the prompts that we used for all the experiments in the paper and hope that's useful.
4. For nudity removal, you are correct that the 1000 images generated in the data path should be safe images. Generating prompts from the llm for "people, men, women" works better. Including "body" in the llm prompting can lead to prompts that lead to unsafe images. Thanks for your suggestion.