Qiankun Liu

Hi @forever-rz , Thanks for your interests. It seems that the second codebook is used for quantization while training transformer. For FFHQ, it will not take so long to get...

Hi, @forever-rz . Sorry for the delayed reply. `keep_ratio` just affects the number of remained pixels in an image, it should not cause such as artifacts. After have a loot...

@forever-rz ， I do not know the number of epochs you have trained on FFHQ and Places2. You can try to visualize the inpainting results of the trained model.

Hi @1997Jessie , Thanks for your interests. Your are right. You need train P-VQVAE firstly, and then train Transformer (with P-VQVAE fixed). Best wishes.

首先需要确认下现有ImageListDataset是否支持你的数据集，不合适的话需要自己写一个dataset，然后修改yaml文件中的dataloader就可以了

感谢关注。代码中根据提供的mask，将白色区域（像素值为255）作为unmask区域，黑色区域（0像素值为0）作为mask区域。你这边的mask应该刚好反过来了。

Hi, @mmaaz60, All the training scripts are uploaded. Note that the appearance network in my paper is the SBC (spatial binary classifier) network in the code.

Hi @zhougeAI , You should provid the path of pvqvae model weights in the configure file of transformer, then it will be automatically loaded. You can refer to this line...

@youyou0805 Thanks for your interest in our project. I do not see any errors in your provided screenshot. But two things are strange. (1) You only provided one pair of...

Hi @boyu-chen-intern , the P-VQVAE is compatible with different image sizes. But the UQ-Transformer is dedicated to the sequence with 1024=32x32 length. Hence, the model can not inpaint images that...