Proteus is by far the best model to use for upscaling, however, its greatest flaw is producing very prominent artifacts when using more extreme values. I have no idea how this stuff works under the hood, but I think this issue could be addressed since it only happens in certain conditions.
The problem at hand are circular dots or patterns that form during processing, these happen when there’s a lot of detail, blurry edges between different colors, or oddly enough sometimes large canvases of solid colors. You can see what I mean and compare the two images below:
This particular frame from Attack on Titan shows some problem areas. Look closely at the clothes of characters and the background by their legs which contains granule detail; I should reiterate that there doesn’t have to be granule detail for this type of artifact to occur, it can appear under completely different circumstances.
I’m optimistic that fine tuning/training the model to compensate for these issues can fix it. 99% of AI upscaling models cannot achieve this level of quality and I can assure this after testing too many to count, it’s the only thing holding Proteus back ATM.
hi, for my case, artemis models are much better than proteus, artemis dehalo and artemis strong dehalo give me the best results, for your video, you have to try to make a forward pass proteus, try artemis aliasing which does not remove details, in the original resolution or in 720p
I’m not completely sure, but I think those dots come from the Sharpen and the Anti-alias/Deblur sliders. When on the Deblur side, those sliders seem to do the same thing.
Auto and Relative to Auto always pick some value too extreme for any DVD I have tried.
The best solution I have found is to use Manual and have those values at 0. You might be able to creep them up, but don’t try more than 3 points at a time, and never move both at the same time.
It stems from higher sharpening values and the other parameters usually just make it worse, setting deblur to 100 actually softens them up a bit without affecting the image quality. I never use auto detection, I have finely tuned settings for anime content that took weeks of experimenting to get the best image quality other than the dots that appear sometimes. I’ve probably gone through hundreds of videos over the past year or so and can spot potential problem areas before upscaling, that’s why I’m asking for more to be done on handling artifacts with higher Proteus values because it can obviously be trained more.
I’ve compared Artemis with my Proteus settings before and can assure that Artemis doesn’t even compare. Antialiasing should only be used for real-life or CG content, it only exaggerates the artifacts I’m experiencing, deblurring is more optimal. Trust me when I say this, I’ve tested about every model and setting in TVAI with denoising/upscaling. The settings I have now are about the best you can get with anime content, the only issue being the artifacts which I think could be fixed with more training on Topaz Labs’ part.
maybe for anime, I work on old 720 x 576 sd videos from 8 mm cassettes. proteus is not bad at sharpening the details but when there are dark passages, it is bad, you have to push the denoise slider to get it clean, whereas with artemis dehalo, the noise is well suppressed and the video comes out less blurry, see images:
On the image from above, artemis dehalo on the left, proteus on the right, proteus keeps a little more detail if well adjusted, but on the image below with the same parameters , we see that proteus removes less noise, and if I increase the denoise or the compression, it removes even more detail than artemis dehalo, so the best thing in this situation is to produce the sunny passages with a certain setting with proteus, and when it gets dark, like the bottom pass, it’s pushing the proteus denoise, so you have to do multiple passes over the same video to do well.