Jianyi Wang

icon indicating a website link https://iceclear.github.io icon indicating an email [email protected]

icon company Nanyang Technological University (NTU) icon location Singapore icon location Ph.D. candidate from MMLab@NTU, mainly focusing on low-level computer vision.

Results 157 comments of Jianyi Wang

Issue training the time-aware encoder

The main problem should be the CPU RAM. You need at least 18G CPU RAM to run the code.

Quality Question

Hi. You may show some examples. We did not have such a problem.

Training and Test datasets

Hi. For training sets, you can refer to [here](https://github.com/xinntao/Real-ESRGAN/blob/master/docs/Training.md#dataset-preparation) For test sets, you may refer to the link in readme.

Training and Test datasets

The original resolution of the test set can reach 2K or larger. The test time can be huge.

Training and Test datasets

We do not finetune on real-world pair datasets. For details you may refer to our paper.

ModuleNotFoundError: No module named 'util_image'

Pls make sure you are running in the correct fold and env.

Output images are in different aspect ratio

Hi. I did not meet such problems. May you share an example?

Output images are in different aspect ratio

I guess the reason is that the input to the unet must be the multiple of 64 since the smallest feature of the unet is 1/64 of the original image...

Output images are in different aspect ratio

I think you can first upscale the input to a resolution slightly larger than the target resolution which is the multiple of 64 and then resize to the target size...

SR image change more vs src image

StableSR is not good at recovering small faces with high fidelity. Actually, such specific scenario is very challenging for general SR. Also, it is not good at recovering Chinese characters...