Chuanxia Zheng

Hi @samuro95, thanks for your interest. Yes, we directly resize images to a larger size and then crop them to 256x256 on all datasets. This is a general data augmentation...

@SherryXTChen. 1. This is a duplicate of https://github.com/lyndonzheng/Pluralistic-Inpainting/issues/4, https://github.com/lyndonzheng/Pluralistic-Inpainting/issues/16, https://github.com/lyndonzheng/Pluralistic-Inpainting/issues/30. 2. We trained the model on CelebA-HQ with centered masks. The training iterations are around 200 epochs. 3. Sure, the...

Hi @liuliuliu11, in the *two_paths* functions, we respectively consider both I_c and I_m in the top and bottom path, where the former is used to enforce the distribution belong a...

Hi @xuwanpeng12, as the generative adversarial learning is sensitive to the network structure, our generation structure follows the work of CycleGAN. At that time, the ResNet performs worse on the...

Hi @kingsleyluoxin , the transforms.Normalize(mean, std) is applied to normalize the input as (x-mean)/std. For the image translation, we normalize the input image from [0,1] to [-1,1].

Hi @ccc870206 This is a duplicate of https://github.com/lyndonzheng/Synthetic2Realistic/issues/35. The pre-trained model is not the original model in the paper. This is my first Ph.D. project and I only stored the...

Hi @ScarletBlaze, you can directly use the coarse results in the first stage for $256\times256$ resolution, which provide high qualitative results for both object removal and foreground completion.

Hi @ChrisChen1023 Thanks for your interesting. I just used the pre-trained models on their websites for the fair comparisons. Due to the different masks and images, it is hard to...

Hi @rose-jinyang The model is able to run on the higher resolution, but we can not ensure the best performance due to the attention model is trained on the 512*512...

Hi @rose-jinyang , thanks for your interest. That's should be ok.