search

ysy31415 / unipaint

Code Implementation of "Uni-paint: A Unified Framework for Multimodal Image Inpainting with Pretrained Diffusion Model"
Apache License 2.0
103 stars 4 forks source link

validation bad cases #4

Open MucciH opened 8 months ago

MucciH commented 8 months ago

Hi! Thanks for sharing the great work! But I might have encountered some problems doing validation. It seems that the model is able to generate good results the first time after finetuning, but the images would turn into bad cases after that. Is there any operation I missed for testing? 如果方便的话想恰个v，My Email address: huangmuqi1@outlook.com Thanks a lot!!!

ysy31415 commented 8 months ago

Hi, if you change either the input image, or mask, or exemplar image, you need to re-run the code to reset the model weights. In other cases, like generating another round of output images, given the same inputs, just with different seed, theoretically this operation should be ok. But in practice I also have the same experience as you (first time is good then bad) which I also feel confused, I just keep trying different seeds after then the expected results will be back. So I think the initial random noise plays a significant role for the final results, generating from random noise is not always stable as expected.

Also, if you have further questions, I am happy to discuss with you via email or wechat. you can send emails to me first: s.y.yang@my.cityu.edu.hk, we can share wechat id then.