I said wowSekiro: Shadows Die Twice. 1080p promotional screenshot. ESRGAN model. 4K downsampled for wallpaper. Cropped, full scale comparison:
Bicubic resampling:
ESRGAN:
With some interpolation between the PSNR model I'm sure it could soften some of the harsher edge enhancement of the arm bands though it's not really apparent when downsampled.
Yeah, messed around with the interp and I've got it in a good spot now.
Here's what you asked for btw.
I tried doing this with dumped Melee textures from Dolphin, but the script have me type conversation errors. Did you run into anything similar?
Now that I've had more time, I've tested ESRGAN, SFTGAN and Waifu2x on a particular Legend of Mana sprite to see how each network performs. On each of them, I magnified the image 4x.
Original Image:
The results can be found here.
My favorite result (ESRGAN trained on Manga109 dataset):
My Observations:
- Before going into this, I expected Waifu2x (urgh, why the name?) to do best because it was trained on illustrations that I think are closest to the image I was trying to upscale.
- However, to my eye, I think the ESRGAN trained on the Manga109 dataset did the best. It had the fewest artifacts and I was very impressed with the result.
- The results from Waifu2x were disappointing - the image just looked as if someone rescaled the original image to 4x using nearest neighbor interpolation and ran a blur filter through it. I tried it on every pre-trained model available and tried combining it with de-noising but frankly everything came out looking very similar to each other.
- The default ESRGAN model did not work so well - lots of artifacts especially in the vine region in the center of the image.
- I ran into a bunch of random issues getting SFTGAN to work. I don't remember all the specifics but there were times it had issues detecting CUDA, crashed my PC, using a lot of resources, etc. I didn't really have time to figure out what was going on so not sure if it's on my end.
- I actually find the artifacts from SFTGAN quite aesthetically pleasing. It looks like oil painting textures.
Don't use Manga model on BR. I think the regular 0.8 mix of the two original models would do better with that game.
Oh yeah. That definitely looks much worse. And you're right, the Manga model upscale looks pretty similar to those original high res renders!
net_PSNR_path = './models/Manga109Attempt.pth'net_ESRGAN_path = './models/Manga109Attempt.pth'net_interp_path = './models/manga109_interp_{:02d}.pth'.format(int(alpha*10))
Can anybody tell me what game this is?Yeah, messed around with the interp and I've got it in a good spot now.
Here's what you asked for btw.
it's a romhack called Super Metroid Ruin
yeah what alexbull_uk said. just mess around with the net_interp.py a bit
Awesome! I'll definitely replay Gothic with the new textures!this Gothic gif got posted on the Gameupscale subreddit. goddamn
Thanks! Didn't know what I was looking for originally but taking it apart this makes it look much more understandable.
git clone https://github.com/xinntao/ESRGANtest.py models/RRDB_ESRGAN_x4.pthpython test.py models/RRDB_ESRGAN_x4.pthpython net_interp.py 0.9 set some "interpolation parameter," where the latter command python test.py models/interp_09.pth generates a super res image by interpolating between the results produced by BOTH the ESRGAN and PSNR models, using the established parameter? Is that right, or am I totally off?python test.py models/RRDB_ESRGAN_x4.pth
python test.py models/RRDB_PSNR_x4.pthcompletely right, although "super res" is all dependent on the original size of the image
yup yup. you can edit the net_interp.py file in a text editor if you want to switch the models with something else like the manga109 one posted earlier in the thread
Right. Makes sense. Just a random thought I had thanks!
