search

Correr-Zhou / MagicTailor

Offical implementation of the paper "MagicTailor: Component-Controllable Personalization in Text-to-Image Diffusion Models".
https://correr-zhou.github.io/MagicTailor/
Apache License 2.0
71 stars 3 forks source link

Demo Code Reproduction #3

Closed AnonymousAchilles closed 1 month ago

AnonymousAchilles commented 1 month ago

Hello,

Thank you for sharing your code! I have a question about replicating the demo code. I followed your instructions for training and using the trained model on the examples of Hinton's face and LeCun's hairstyle. However, the background didn't generate a beach scene as expected from the prompt: " with , on the beach." Could there be a reason for this discrepancy?

result

Correr-Zhou commented 1 month ago

It looks fine here when I run the original code. I will go check the released code. Before that, you can try adjusting the hyperparameters for better results. Thanks for your feedback.

pic1

Correr-Zhou commented 1 month ago

Hi, I have updated inference.py, aligning the inference settings with the original ones. The generation quality is back to normal now. Thanks for your feedback on this issue. : ) result