search

yardenfren1996 / B-LoRA

Implicit Style-Content Separation using B-LoRA
MIT License
307 stars 22 forks source link

Curious and splitted results #20

Closed clementjambon closed 4 months ago

clementjambon commented 4 months ago

I tried to play with your pre-trained LoRA weights (using directly https://github.com/yardenfren1996/B-LoRA/blob/main/B_LoRA_inference.ipynb) and I get very curious results.

For "Image Stylization Based On Text Reference", when using the prompt "A [v3] made of gold", things look fine (although not similar). However, using other materials such as "A [v3] made of wood" yields extremely weird results like this: image (19)

When running exactly the example for "Consistent Style Generation", I get splitted results too image (20)

yardenfren1996 commented 4 months ago

Hi, @clementjambon

Thanks for reaching out.

We are aware that some "bad seeds" may yield weird results when using our method. In your example, seed 48 did not work well with the prompt "A [v3] made of wood." To investigate further, I randomly selected another 9 seeds, generated 10 results in total, and plotted them in the following image:

Screenshot 2024-07-24 at 13 09 48

As you can see, our method is consistent with 8 out of 10 different seeds. (except 48, 668)

We are confident in our results, but sometimes using a different seed or adjusting the alpha can improve them even further. Thank you.