search

tencent-ailab / PCDMs

Implementation codeļ¼šAdvancing Pose-Guided Image Synthesis with Progressive Conditional Diffusion Models
Apache License 2.0
150 stars 9 forks source link

It does not work with custom poses or different poses other than the pair given in the example. #21

Open adityapandey9 opened 4 months ago

adityapandey9 commented 4 months ago

It does not work with custom poses or poses other than the pair in the example.

I have used the image img3.png given in the images folder with pose2, I am getting the below output.

Screenshot 2024-07-03 at 1 09 10 PM

When I try to use a custom pose like the one below with any subject images, it does not work.

Screenshot 2024-07-03 at 1 11 00 PM

Any help would be appreciated.

muzishen commented 4 months ago

Thanks for your attention, the first case that fails may require multiple sample generation several times. Because we have limited data and computing power, but the stabilizing effect outside the domain requires more data and training steps. The problem in the second case seems to be that the posture is different from the one we used. It is suggested to use the posture detection tool we provided.

adityapandey9 commented 4 months ago

Hi @muzishen thank you for your response. So, do I need to retrain the model for each new subject? This is similar to the dreambooth, right?

muzishen commented 3 months ago

Hi @muzishen thank you for your response. So, do I need to retrain the model for each new subject? This is similar to the dreambooth, right?

Don't need it. As long as the number of ids is sufficient, then our ckpt can continue training.