Xu Yu

Results 15 comments of Xu Yu

> Hi, yes this should not take too long. Can you specify what are the flags that you are using when computing the FID? > > -Gaurav hi, I have...

re-install the evironment works for me, maybe try to re-use `pip install -e .`

you put your custom data here. Which is 'Input images' in Figure 1 of the paper.

@AkiKagura Hi, I would like to ask about face idnetity preservation. Does the results and the training data look like the same person? According to my experiments, the results and...

@AkiKagura thanks for your reply, I set training steps to 2000, and find it's good to generate man portrait(or face reconstruction). However, face is distorted as you mentioned above if...

@AkiKagura Maybe it's just inherent limitation of DreamBooth? Not modeling or optimizing on face with rich background. However, I find results in [this video](https://www.youtube.com/watch?v=Z-hyKADmHmE) seems satisfied. But have not tested...

@AkiKagura If you generate imgs of half-body, distorted faces can be avoided, however the face identity is far from training set, be like ![huangbo](https://user-images.githubusercontent.com/94051337/203467822-b3d06950-4c0b-4b59-b468-9c1458dea1e0.PNG) training set and prompt is same...

@howardgriffin hi, meet the same question, I'd like to ask any idea to solve it? Really thanks!

@1289600760 hi, really thanks for you suggestion, and have you finally find the reason?

hi, I use: ``` model_type = Model_Type.SDXL scheduler_type = Scheduler_Type.DDIM ``` I didn't modify anything in ` inversion_example_sdxl.py`.