ESRGAN is developed by
Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, Chen Change Loy
Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, Chen Change Loy
2019 (outdated)
____________________________________________________________________
I did many tests and those are my results and experiences.
You can download the models at the bottom of the page!
Recommended Hardware: Video Card with 8GB and higher!
Goal: I tried to upscale low-res pixel art images to the highest quality without retouching. I trained many new models to achieve the results.
Conclusion: More interpolation is used, closer the results are to your trained images. Less interpolation means more regular PSNR, which makes it smoother and better for low-res images. More interpolation means it will get closer to images you provided for training. (it should, at least)
Zoom-in to see full resolution animated GIF images
ESRGAN calculations automatically turns off at 500 000 interpolations.
HOW to prepare images for training?
1. Find high quality images that you wanna use for training.
2. Crop them by using python script: ...codes\scripts\extract_subimgs_single.py (change the input_folder to where your images are, set crop_sz and step to 128.) Run it.
3. Downscale cropped images to 4x of the size, (128/32 = 8GB VRAM, 256/64 = 20GB VRAM)
(important is the method of downscaling NEAREST NEIGHBOUR
if you use any other like; Bicubic, B-spline, Mitchell...you will fail!)
RESULTS (animated gifs):
____________________________________________________________________
Training images examples:
Indy3
(full size 2560x512 up-scaled)
Monkey Island 1
(full size 1280x800 up-scaled)
Monkey Island 2
(full size 1280x576 up-scaled)
Sam & Max
(full size 1280x800 up-scaled)
IJ Last Crusade & IJ Fate of Atlantis
(full size 1500x380 up-scaled)
Beneath a steel Sky & Day of the Tentacle
(full size 1280x800 up-scaled)
IJ Fate of Atlantis
(full size 1280x800 up-scaled)
Sam & Max
(full size 1280x800 up-scaled)
2019
You can download the models at the bottom of the page!
Recommended Hardware: Video Card with 8GB and higher!
Tested on 1080Ti GTX 11GB
Tested on 1070 GTX 8GB
(approx. 2 minutes per 100 interpolations)
Tested on 1070 GTX 8GB
(approx. 2 minutes per 100 interpolations)
Goal: I tried to upscale low-res pixel art images to the highest quality without retouching. I trained many new models to achieve the results.
Conclusion: More interpolation is used, closer the results are to your trained images. Less interpolation means more regular PSNR, which makes it smoother and better for low-res images. More interpolation means it will get closer to images you provided for training. (it should, at least)
Zoom-in to see full resolution animated GIF images
ESRGAN calculations automatically turns off at 500 000 interpolations.
HOW to prepare images for training?
1. Find high quality images that you wanna use for training.
2. Crop them by using python script: ...codes\scripts\extract_subimgs_single.py (change the input_folder to where your images are, set crop_sz and step to 128.) Run it.
3. Downscale cropped images to 4x of the size, (128/32 = 8GB VRAM, 256/64 = 20GB VRAM)
(important is the method of downscaling NEAREST NEIGHBOUR
if you use any other like; Bicubic, B-spline, Mitchell...you will fail!)
RESULTS (animated gifs):
____________________________________________________________________
DONALD DUCK COMIC BOOK
5000 - 110 000 interpolation
HR - 256pixels / LR - 64pixels
(3 days of calculations)
110 Images total - Bicubic
110 Images total - Bicubic
Training images examples:
Indy3
(full size 2560x512 up-scaled)
Monkey Island 1
(full size 1280x800 up-scaled)
Monkey Island 2
(full size 1280x576 up-scaled)
____________________________________________________________________
CARTOON PAINTED
5000 - 80 000 - 400 000 interpolation
5000 - 80 000 - 400 000 interpolation
HR - 128pixels / LR - 32pixels
(7 days of calculations)
4669 images total - Nearest
4669 images total - Nearest
Training images examples:
Sam & Max
(full size 1280x800 up-scaled)
IJ Last Crusade & IJ Fate of Atlantis
(full size 1500x380 up-scaled)
Beneath a steel Sky & Day of the Tentacle
(full size 1280x800 up-scaled)
DISNEY MOVIE PINOCCHIO
5000 - 35 000 interpolation
HR - 128pixels / LR - 32pixels
(6 hours of calculations)
2409 Images total - Nearest
2409 Images total - Nearest
Results:
Monkey Island 1
(full size 1280x576 up-scaled)IJ Fate of Atlantis
(full size 1280x800 up-scaled)
Sam & Max
(full size 1280x800 up-scaled)
____________________________________________________________________
DISNEY MOVIE HERCULES
5000 - 40 000 interpolation
HR - 128pixels / LR - 32pixels
(1 days of calculations)
5557 Images total - Nearest
Training images examples:
20528 Images total - Nearest
Training images examples:
5557 Images total - Nearest
Training images examples:
Adventure Games
____________________________________________________________________
DISNEY MOVIE LION KING
5000 - 60 000 interpolation
HR - 128pixels / LR - 32pixels
(2 days of calculations)20528 Images total - Nearest
Training images examples:
Results:
IJ Fate of Atlantis
(full size 1280x576 up-scaled)
Loom
(full size 1920x576 up-scaled)
Monkey Island 1
(full size 1280x800 up-scaled)
6852 Images total - Nearest
Training images examples:
Loom
(full size 1920x576 up-scaled)
Monkey Island 1
(full size 1280x800 up-scaled)
Monkey Island 2
(full size 1500x342 up-scaled)
____________________________________________________________________
MANGA109 my own recal
5000 - 250 000 interpolation
HR - 128pixels / LR - 32pixels
(5 days of calculations)6852 Images total - Nearest
Training images examples:
Results:
Adventure Games
____________________________________________________________________
MANGA109 kingdomakrillic
(interp? no information provided)
HR - 128pixels / LR - 32pixels
(6 hours with 1050 Ti GTX) - Nearest Neighbour
Results:
Adventure Games
(full size 1K 2K 3K up-scaled)
____________________________________________________________________
RRDB PSNR VS RRDB ESRGAN
Blending RRDB_PSNR with RRDB_ESRGAN models with alpha 0.1-1.0
Adventure Games
(full size 1280x800 up-scaled)
____________________________________________________________________
WAIFU2x
Image Super-Resolution for Anime-style art using
Deep Convolutional Neural Networks.
Deep Convolutional Neural Networks.
Adventure Games
____________________________________________________________________
LET'S ENHANCE.io
It's SRGAN, and the results are superior to SRGAN-tensorflow’s defaults.
Adventure Games
(full size 1280pixels to 2500pixels up-scaled)2019