search

yardenfren1996 / B-LoRA

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

Not always true. #5

Open xingp-ng opened 7 months ago

xingp-ng commented 7 months ago

We believe that we have reproduced the appropriate results, but there are still some questions that we would like to be answered.

  1. The results are generated in such a way that it usually takes one out of 8 to find a suitable result, making it difficult to use them in practice.

  2. These results may not be aligned with the content image.

  3. Is there any technique to alleviate these problems？

yardenfren1996 commented 7 months ago

Regarding your questions, since it is an optimization process. There's a possibility that training different LoRAs on the same image will yield different results due to different initializations. Sometimes, the optimization process struggles to 'learn' the given concept perfectly. I recommend training with different seeds, or adjusting other training parameters. And since it is a personalization technique, the resulting image may not align perfectly with the content image. I suggest integrating our approach with content preservation techniques like ControlNet, although I haven't personally tested this.

liusida commented 5 months ago

Ha-ha, interesting workaround. 5 B-LoRAs with different seeds have the comparable size of a LoRA model xD