versatile diffusion reconstruction is worse than reported about brain-diffuser HOT 7 CLOSED

@cnwangy Well, I never encountered any results like these but it seems the diffusion model doesn't seem to working properly, have you changed any parameters in the versatilediffusion_reconstruct_images.py like strength or mixing?

from brain-diffuser.

I also followed the steps you mentioned for training, and the training results were the same as his, very blurry, like partial screenshots of the original image.

from brain-diffuser.

I think I have solved the problem. It is the version of the transformers lib. When I use v4.33.2 the extracted clipvision features are different from v4.19.2. Using the latter one I can replicate your results. Thanks a lot!

from brain-diffuser.

I think I have solved the problem. It is the version of the transformers lib. When I use v4.33.2 the extracted clipvision features are different from v4.19.2. Using the latter one I can replicate your results. Thanks a lot!

May I ask how you solved it and how to use v4.33.2? Can you provide me with a detailed explanation? Because I am a novice, please forgive me. Thank you.

from brain-diffuser.

I think I have solved the problem. It is the version of the transformers lib. When I use v4.33.2 the extracted clipvision features are different from v4.19.2. Using the latter one I can replicate your results. Thanks a lot!

May I ask how you solved it and how to use v4.33.2? Can you provide me with a detailed explanation? Because I am a novice, please forgive me. Thank you.

You can just follow the README to use transformers==v4.19.2 as in the environment.yaml. Do not upgrade to v4.33.2.

from brain-diffuser.

I think I have solved the problem. It is the version of the transformers lib. When I use v4.33.2 the extracted clipvision features are different from v4.19.2. Using the latter one I can replicate your results. Thanks a lot!

May I ask how you solved it and how to use v4.33.2? Can you provide me with a detailed explanation? Because I am a novice, please forgive me. Thank you.

You can just follow the README to use transformers==v4.19.2 as in the environment.yaml. Do not upgrade to v4.33.2.

Thank you very much for your guidance. I have successfully reproduced it.

from brain-diffuser.

I think I have solved the problem. It is the version of the transformers lib. When I use v4.33.2 the extracted clipvision features are different from v4.19.2. Using the latter one I can replicate your results. Thanks a lot!

May I ask how you solved it and how to use v4.33.2? Can you provide me with a detailed explanation? Because I am a novice, please forgive me. Thank you.

You can just follow the README to use transformers==v4.19.2 as in the environment.yaml. Do not upgrade to v4.33.2.

Thank you very much for the solution you provided. However, I encountered the following error after downgrading the transforms library.
'(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /openai/clip-vit-large-patch14/resolve/main/vocab.json (Caused by ConnectTimeoutError(<urllib3.connection.HTTPSConnection object at 0x7f1af21956d0>, 'Connection to huggingface.co timed out. (connect timeout=10)'))"), '(Request ID: 3df39c48-06cc-45e7-8096-22b9ca4c9f91)')' thrown while requesting HEAD https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/**vocab.json**
I've tried to resolve it, but it seems that Hugging Face no longer provides older versions of the models. I'm not sure if you've encountered the same issue or if you have any suggestions for a solution. Once again, thank you for your help

from brain-diffuser.