search

xuekt98 / BBDM

BBDM: Image-to-image Translation with Brownian Bridge Diffusion Models
MIT License
230 stars 24 forks source link

Question about image_size in 'Template-BBDM'.yaml #54

Open HaHaFish1014 opened 3 weeks ago

HaHaFish1014 commented 3 weeks ago

Thank you for your great work and share your code! I have a question about the image_size in 'Template-BBDM'.yaml. When i use the 'Template-LBBDM-f8'.yaml to train in my own datasets for colorization, i found that the results of sample_to_eval are very strange. I solved it by using the 'Template-BBDM'.yaml. However, using 'Template-BBDM'.yaml only generate 6464 images. I want to generate 128128 and 256*256 images. What should i do to modify the yaml setting? Can i just simply modify data:image_size and Unet:image_size? Hope for your reply.

sumit171993 commented 3 weeks ago

Hi @HaHaFish1014 , I am unable to access the pre-trained model CelebHQ since I am unable to make my account on the Baidu link since I am from India; please provide the pre-trained weights that you are using for fine-tuning your dataset; if possible by some google drive link if you are doing so. Otherwise, I need to train from scratch, but I have limited images right now for training, i.e., 2101 only. Regarding your question, I am using 256 x256 images for training, and for Unet using 64 x64, I think we should not change the Unet dimensions. Thanks and Regards, Sumit, This is my emailID: ee20d042@smail.iitm.ac.in

HaHaFish1014 commented 2 weeks ago

Hi @HaHaFish1014 , I am unable to access the pre-trained model CelebHQ since I am unable to make my account on the Baidu link since I am from India; please provide the pre-trained weights that you are using for fine-tuning your dataset; if possible by some google drive link if you are doing so. Otherwise, I need to train from scratch, but I have limited images right now for training, i.e., 2101 only. Regarding your question, I am using 256 x256 images for training, and for Unet using 64 x64, I think we should not change the Unet dimensions. Thanks and Regards, Sumit, This is my emailID: ee20d042@smail.iitm.ac.in

Thank you for your advice! I did not use the pre-trained weights. I used my own dataset to train from beginning. About pre-trained weights, you should ask the author because i think i do not have the authority to disseminate the pre-trained weights. May i ask you if you use the 'Template-BBDM.yaml' to train? How did you use 256256 images for training? I cant train if i use 256256 images as input. Only modify the 'Template-BBDM.yaml' setting can solve it, but i am worried that it will break the proper function of the attention layers in Unet.