I've heard good thing about Topaz but I guess one would have to test it on anime images
Based on your sample (and your sample only as I only have Waifu2x on my computer), I think Topez's is actually not bad.
I would prefer Topez's over waifu2x's "oily (overfiltered feeling)" result if not for the artifact introduced by JPEG - which Waifu2x definitely has a lead on removing them.
Yes, the Waifu2x one is more "clean" but also it also loses lots of details (see the texture on her shoulder etc.).
I wonder if the source file is png, it may have even clearly better results than Waifu2x.
I would prefer Topez's over waifu2x's "oily (overfiltered feeling)" result if not for the artifact introduced by JPEG - which Waifu2x definitely has a lead on removing them.
Yes, the Waifu2x one is more "clean" but also it also loses lots of details (see the texture on her shoulder etc.).
I wonder if the source file is png, it may have even clearly better results than Waifu2x.
I'm pretty sure those are not "details" on her shoulder, they're just noise/artifacts, but I will do another test with very clean images and pure upscaling.fireattack said:
Based on your sample (and your sample only as I only have Waifu2x on my computer), I think Topez's is actually not bad.
I would prefer Topez's over waifu2x's "oily (overfiltered feeling)" result if not for the artifact introduced by JPEG - which Waifu2x definitely has a lead on removing them.
Yes, the Waifu2x one is more "clean" but also it also loses lots of details (see the texture on her shoulder etc.).
I wonder if the source file is png, it may have even clearly better results than Waifu2x.
Here's the result:
https://ibb.co/ZmVzgBk
Topaz Labs and Photoshop somehow re-introduces jpeg artifacts into the image where none existed....
I also just tested straight up photograph, and there really isn't much of a difference. Gigapixel looks sharper but if you just sharpen the other two they basically look identical.
https://ibb.co/kmkjP5b
I insist that Gigapixel is just a scam, like many of their other filters.
Would be easier to tell if you can provide the original image (I can't find it via reverse image search), but yeah if that's the case, then these "details" are no good.
The waifu2x one still lost substantial details though (for example, the end of her eyelash on the right eye that extended to orange hair is totally gone), but that's more about the denoise pre-processing. That's why I normally won't use its denoise filter (but it's kinda unavoidable in order to get acceptable result if the source is noisy, admittedly).
> somehow re-introduces jpeg artifacts into the image
More like some haloing-like artifact due to the upscaling process, kinda common in traditional algorithms.
The waifu2x one still lost substantial details though (for example, the end of her eyelash on the right eye that extended to orange hair is totally gone), but that's more about the denoise pre-processing. That's why I normally won't use its denoise filter (but it's kinda unavoidable in order to get acceptable result if the source is noisy, admittedly).
> somehow re-introduces jpeg artifacts into the image
More like some haloing-like artifact due to the upscaling process, kinda common in traditional algorithms.
https://yande.re/post/show/515674fireattack said:
Would be easier to tell if you can provide the original image (I can't find it via reverse image search), but yeah if that's the case, then these "details" are no good.
The waifu2x one still lost substantial details though (for example, the end of her eyelash on the right eye that extended to orange hair is totally gone), but that's more about the denoise pre-processing. That's why I normally won't use its denoise filter (but it's kinda unavoidable in order to get acceptable result if the source is noisy, admittedly).
> somehow re-introduces jpeg artifacts into the image
More like some haloing-like artifact due to the upscaling process, kinda common in traditional algorithms.
I see the eyelash thing, seems pretty minor to me, I ran the same image with less denoise strength, which does keep the eyelash but also kept some of the noise. I picked this image in the first place because it is the kind that I most desperately need good upscaler for, fantastic art but low res and jpeg only.
BTW the other algorithms were ran at the highest denoise settings.
waifu2x has been ported to the iPad/iPhone. I like it so far and it will save as PNG. 2X and 4X upscale.
https://itunes.apple.com/us/app/waifu2x/id1286485858?mt=8
https://itunes.apple.com/us/app/waifu2x/id1286485858?mt=8
What are your suggested settings and models? I'd like to give it a try.chen said:
https://yande.re/post/show/515674
I see the eyelash thing, seems pretty minor to me, I ran the same image with less denoise strength, which does keep the eyelash but also kept some of the noise. I picked this image in the first place because it is the kind that I most desperately need good upscaler for, fantastic art but low res and jpeg only.
BTW the other algorithms were ran at the highest denoise settings.
I'm hoping https://github.com/imxieyi/waifu2x-mac will be published at some point.
I just built it on my 2012 Macbook running 10.13.6Radioactive said:
I'm hoping https://github.com/imxieyi/waifu2x-mac will be published at some point.
I had to download Xcode 10, install it and then compile the app as instructed.
Anyone can do this.
It works, but the Preferences are dimmed and it's 2X only. Maybe because it's not ready or there's something I don't know.
Results are here: https://ibb.co/hcZGjXy It does output the file as png only. 15.8 seconds conversion time on my old macbook pro.
Original here: https://twitter.com/Sohuzi/status/1065258726154657794
Looking at Gambier Bay's antenna hair ornament, the online waifu2x edges it out slightly in sharpness. But it's pretty decent and will enlarge the files that are over the online version limits.
So if you want to play with it until it's officially released, you can have waifu2x on MacOS. 10.13 and 10.14 only.
Thanks Genex. I will wait until it's officially released if you are seeing those kind of issues.
Ok, and here is another sample: https://ibb.co/F8Wcs1J
I've built it on 10.13 and now I'll try it on 10.14.
Some asides: If you subscribe to Adobe's Creative Cloud Photoshop, avoid upgrading to 10.14 for now.
Also Apple is going to drop support for 32 bit Intel-based apps in 10.15 So if you have any ancient copy of CS3-CS6, don't go to 10.15.
I've built it on 10.13 and now I'll try it on 10.14.
Some asides: If you subscribe to Adobe's Creative Cloud Photoshop, avoid upgrading to 10.14 for now.
Also Apple is going to drop support for 32 bit Intel-based apps in 10.15 So if you have any ancient copy of CS3-CS6, don't go to 10.15.
Was anything done with the sample? It suffers a lot with artifacting.
Let's try this: https://imgur.com/a/BsNEABX and see if it's the image uploader.
Also compiled Waif2x for my Mac 10.14 MBP, again successfully, but still dimmed preferences.
Also compiled Waif2x for my Mac 10.14 MBP, again successfully, but still dimmed preferences.
That's better. Was this filtering and/or upscaling?
All waif2x-mac does is 2X upscaling. There are three buttons:
Pick: Open a file in the left window.
Process: Upscale the file. When that is done, you see the elapsed time for the conversion and the processed file is in the right window.
Save: Name the file from the generic "untitled", pick a directory, and save.
Since preferences are dimmed out, I have no idea what level of filtering was used.
The Mojave build is the same way on my other Macbook Pro.
So that's all it offers at this point. I hope to see if I can get ahold of the author. He also writes waif2x-ios which is in the iphone/ipad app store. He warns he hasn't published waifu2x-mac yet, and any similar apps offered for sale are not his.
I will say it's faster than waifu2x-caffe on my windows machine.
He does provide a Command line version with the command options if you don't mind that.:
Usage: waifu2x [options]
-t, --type:
Image type - a for anime (default), p for photo
-s, --scale:
Scale factor (1 or 2)
-n, --noise:
Denoise level (0-4)
-i, --input:
Input image file (any format as long as NSImage loads)
-o, --output:
Output image file (png)
-h, --help:
Print usage
Pick: Open a file in the left window.
Process: Upscale the file. When that is done, you see the elapsed time for the conversion and the processed file is in the right window.
Save: Name the file from the generic "untitled", pick a directory, and save.
Since preferences are dimmed out, I have no idea what level of filtering was used.
The Mojave build is the same way on my other Macbook Pro.
So that's all it offers at this point. I hope to see if I can get ahold of the author. He also writes waif2x-ios which is in the iphone/ipad app store. He warns he hasn't published waifu2x-mac yet, and any similar apps offered for sale are not his.
I will say it's faster than waifu2x-caffe on my windows machine.
He does provide a Command line version with the command options if you don't mind that.:
Usage: waifu2x [options]
-t, --type:
Image type - a for anime (default), p for photo
-s, --scale:
Scale factor (1 or 2)
-n, --noise:
Denoise level (0-4)
-i, --input:
Input image file (any format as long as NSImage loads)
-o, --output:
Output image file (png)
-h, --help:
Print usage
Probably model related, or simply GPU acceleration if you aren't using CUDA on windows.Genex said:
I will say it's faster than waifu2x-caffe on my windows machine.
In terms of speed UpRGB > RGB > CUnet > UpResNet10 > TTA modes
It appears waifu2x-mac defaults to UpRGB, while current versions of waifu2x-caffe and waifu2x web default to CUnet which is significantly slower yet offers better quality.
How would you get the Mac GUI version to use CUNet if you wanted it?
It could be as simple as converting the CUnet models (t7 or json) to coreml (mlmodel) format and replacing files.
If you wanted to convert the models yourself rather than request the developer do so, it looks like one of these tools may do it:
https://apple.github.io/coremltools
https://github.com/Microsoft/MMdnn
https://github.com/prisma-ai/torch2coreml
If you wanted to convert the models yourself rather than request the developer do so, it looks like one of these tools may do it:
https://apple.github.io/coremltools
https://github.com/Microsoft/MMdnn
https://github.com/prisma-ai/torch2coreml
chen
In case anyone is wondering, Waifu2X is still king
Well I tried and I think the answer is actually pretty definitive, Waifr2X with the right model basically is producing perfect results, it's as if someone manually re-drawn the same image in higher resolution.
https://ibb.co/gzpwqHd
Topaz Labs Gigapixel is clearly designed for photos and not illustration, it likes to add detail that shouldn't be there and sometimes it confuses JPEG artifacts as detail.
Photoshop's own solution is not bad but it's just not as good as Waifu2X. at least it runs super fast compared to the rest.
BTW I also compared Photoshop versus Gigapixel on real life photos, there's almost no difference between the two, Gigapixel is like Photoshop but with a sharpening filter, it's not worth the money.