search

adobe-research / custom-diffusion

Custom Diffusion: Multi-Concept Customization of Text-to-Image Diffusion (CVPR 2023)
https://www.cs.cmu.edu/~custom-diffusion
Other
1.84k stars 135 forks source link

if it OK with no regularization? my results don't have any consistancy with reference images. #85

Open taehyeonEum opened 8 months ago

taehyeonEum commented 8 months ago

Before addressing my problem, Thanks for great works you've done and all kindness.

I just did fine-tuning and inference for multi-object cat and wooden pot through diffusers. (WITH NO REGULARIZATION) But, when I check the generated images. There was no Identity consistancy with reference images when used in training phase, both cat and wooden pot.

I wonder this loss of consistancy due to absence of regularization. Or other problem.. I'm looking forward for your comment..!

you can see codes in my github https://github.com/taehyeonEum/CD_not_official.git

Thank you

taehyeonEum commented 8 months ago

May be I should try more..!! After trying more If I still have the problem I will ask again..!!

ZZZBBBZZZ commented 8 months ago

The url "https://knn.laion.ai/knn-service" always returns me [520],what can i do? Can i train with no regularization?

taehyeonEum commented 7 months ago

I tried with no regularization but performance was bad.... So I tried with normal webcrawling technique. And It works fine.

If you want check my codes. https://github.com/taehyeonEum/CustomDiffusion_SceneRegeneration_KT_internship.git