Thank you for sharing your great work!

I want to run test.py using your pretrained models. Do I need to prepare the mask images and file lists by myself? If you have the code to generate them, would you mind sharing it?

Hi, thank you for your great work! When training the vqvae network, the training time is very long, could you tell me how to accelerate? Thanks for replying!

Hi! Thanks for your great work.

I wonder about the effect of discrete textural features. It seems that these discrete textural features do not enroll in the image generation part. Are the benefits to splits structural and textural features are better disentanglement?

I am a newcomer in this field and am very interested in your program. Can you tell me how filelist is generated? Can you publish your code? Thank you.

I use a data set with picture sizes of 256*256 and use the pre-trained model you gave for testing. The program reports such an error. Do you know what I did wrong? Thank you.

Could you please elaborate a little on input image size and size of mask for respective datasets? like celeb input image is 266x266 and the mask is written 256x256, which seems to be wrong.

Specifically, could you share the input image size places 2, and the size of the mask?

In training: https://github.com/USTC-JialunPeng/Diverse-Structure-Inpainting/blob/main/README.md#:~:text=Modify%20data/data_loader.py%20according%20to%20the%20dataset.%20For%20CelebA-HQ%2C%20we%20resize%20each%20image%20to%20266x266%20and%20randomly%20crop%20a%20256x256.%20For%20Places2%20and%20ImageNet%2C%20we%20randomly%20crop%20a%20256x256

In testing https://github.com/USTC-JialunPeng/Diverse-Structure-Inpainting/blob/main/README.md#:~:text=Collect%20the%20testing%20set.%20For%20CelebA-HQ%2C%20we%20resize%20each%20image%20to%20256x256.%20For%20Places2%20and%20ImageNet%2C%20we%20crop%20a%20center%20256x256.

I have tested the pre-trained model and found the results to be very off, attaching two sample-generated images.

Hi and first of all thanks for the great code!

Can I ask why you stack a matrix of ones in the penultimate channel here? https://github.com/USTC-JialunPeng/Diverse-Structure-Inpainting/blob/467b7a9ea3b783c09ec5beb941085bd8e75de2d3/net/structure_generator.py#L30

Wouldn't it make more sense to have x = tf.concat([x, ones_x*(1-mask), ones_x*mask], 3) ? (i.e. you concat the mask as its one-hot encoding).

Best, P.

I am wondering if there is a discrepancy between the how inference is explained in the paper and how it is implemented in test.py. To be specific, the paper says "During inference, only Gs and Gt are used." but the VQ-VAE model is loaded in test.py and used to do inference (see here and here for example).

I am new to tensorflow and not super familiar with VQ-VAE, so I might be missing something, but I think since the VQ-VAE encoder is getting the full image as input (see here) there might information leak from the full image to the inpainting module at inference time. Please correct me if I'm wrong.

Thank you.

Hi, you mention in the README : "For Places2 and ImageNet, we crop a center 256x256." It seems that in your test.py code, you resize all images to 256x256 and don't do any cropping. What did you do for your results in the paper ?

Hello, thank you for your amazing implementation Can you provide a demo file or a colab notebook for custom images, it will be much appreciated and it'll help everyone,

Thank you

Hi! I am using the model for generating new faces from masked faces. I have run test.py and modify the code a little for removing face mask.

The model is not performing on CelebA masked faces , kindly guide for training for masked face, do i need to pass masks ?

hello,when I use your code training, it occurs indexError: in this part: for i in range(4): gt_i = ((gt[i] + 1.) * 127.5).astype(np.uint8) masked_i = ((masked[i] + 1.) * 127.5).astype(np.uint8) complete_i = ((complete[i] + 1.) * 127.5).astype(np.uint8) recons_gt_i = ((recons_gt[i] + 1.) * 127.5).astype(np.uint8)

train_structure_generator.py： Traceback (most recent call last): File "train_structure_generator.py", line 366, in nn.structure_visual(gt_np, masked_np, recons_gen_np, recons_gt_np, (i + 1), args.image_size, folder_path) File "D:\pythonProject\5_19\VQ_VAE\net\nn.py", line 173, in structure_visual gt_i = ((gt[i] + 1.) * 127.5).astype(np.uint8) IndexError: index 1 is out of bounds for axis 0 with size 1

How many dimensions is the parameter gt? What data parameter is gt?thank you

Training structure_generator.In order to solve the problem of insufficient video memory during generator, reducing batch_size leads to an error when saving the training results : Index 1 is out of bounds for axis 0 with size 1.I hope I can get your help. Thank you!

The paper is exciting! Thanks to the authors. The program runs perfectly on the GPUs(two 2080), but the same version(TensorFlow-gpu=1.12.0 cuda=9.0 cudnn=7.6.5) cannot be adapted to the two A4000. Does anyone have a solution to the version problem?