mlpc-ucsd / patch-dm Goto Github PK

Hi, I have a question regarding the training on natural image dataset. How long did you guys trained on the natural images dataset? Thank you!

Thanks to your great work! I did experiment on LSUN Church dataset.Following on your paper, global conditions of this dataset is provided by a U-Net's encoder. After that, train another latent diffusion model for unconditional image synthesis by the pretrained model. But the loading model code is for your own landscape images not for low resolution images.I wonder how to train LSUN Church dataset on latent diffusion model?Sincerely looking forward to your reply.

I wonder if I can train on Patch-DM and then use the pretrained weights on DDIM models.

Hi Zheng and Mengqi,

Thank you for your amazing work and making your code public! I wonder if you could kindly provide some insights on my experiments below?

I am training on the truncated CelebA dataset, which only has 5000 256x256 images. And I am using CLIP embedding. My batch-size is 96.

1. How many epoch would it take for the train.py to converge?
2. How many epoch would it take for the train_latent.py to converge?
3. Should I run train.py longer or train_latent.py longer when the generated images have grid-like artifacts?

I try to run train_latent on my custom dataset but I run into NameError: free variable 'patch_num_x' referenced before assignment in enclosing scope

Thank you so much for sharing the amazing work! I have a question regarding training image resolution, in img2lmdb.py I only see it converting size up to 1024. Do you support even higher resolutions such as 2k or 4k and patchify? My dataset mostly 3k resolution images. Thank you!

Is data_path in initialize.py the same as lmdb_path in img2lmdb.py? I use lmdb_path in img2lmdb.py as data_path in initialize.py. The error message is as follows. After a long time of debugging, the error still occurs.Could you please help me answer my question? Thank you!

PIL.UnidentifiedImageError: cannot identify image file <_io.BytesIO object at 0x7f642ed3a5e0>

Hello, thank you for presenting your work.I have a question about border artifacts during training.

• My understanding is the grid-like artifact will gradually gets better during training iterations, but still remain obvious in early training stage, is that the case during your expreiment?
• And do you think it helps to constrain the consistency between feature collagre output (output shift) and image collage output (otuput nonshift)

Thanks for your great work! But I have some questions about code. Is the BeatGANsAutoencModel in unet_autoenc.py counterpart with the encoder of UNet? But the tensors of pos_emb and pos_emb_new seem not to be used in the model. Does it not use the position embedding or I misunderstand? Sincerely looking forward to your answer!