Open giovanlee opened 3 days ago
Hi @giovanlee. Thank you for reaching out!
We used CompVis stable diffusion 1.4 for our experiments, which was originally fine-tuned on images of size 512*512. Another reason was that our ControlNet could only be trained on images of lengths 512 and 1024 (according to the HF implementation). So, to have a fair comparison with ControlNet, we needed to keep the resolution at 512.
I hope this answers your question. Let me know if you have any other questions. Best, Pooria
Hi.
First of all. Thanks for your interesting research. I am interesting in finetuning stabe diffusion with my own medical dataset. However, I have some a question for you CAMUS dataset has image size of 384, 384. However you finetuned stable diffusion with --resolution=512. Is there any reason for this training ?
Thank you