![]() See what the models in the ESRGAN database can do: I'd follow a similar process for non-anime. ![]() Even with all that stuff applied and tuned the final result won't get close to a good NN great work, but what about upscaling non-anime stuff? Thanks. WarpSharp may distort the geometry of objects like rectangles and a silhouette on a dark background. ![]() Sharpening by filters at some point creates artefacts and increases noise. These simple filters can work to a point but aren't in the same league as a good neural net. I've tried most sharpening filters that come with Hybrid, including the line darkening stuff. Lastly to those who think upscaling by neural networks is nothing but sharpening. Knowing that I reduced halos of the oversharpened source just halfway in pre-processing with Hybrid:Ĭlip = havsfunc.DeHalo_alpha(clip, rx=2.2, ry=2.2, darkstr=0.25, brightstr=0.50) The model reduces halo artefacts and noise. The model seems to have added a very subtle aquarelle paper texture that adds to the artistic feel but probably won't survive final h264 compression. Anime lines are a bit thinner and clear, everything is as crisp as it gets considering the source. Recovered by the model in a phenomenal way. Look at the black lines in A-ko's hair that were very faint and almost gone. I attached some examples (with preliminary color correction for the Cupscale images). For now, I'll take that compromise as well. I wish there were a 2x version of this model because the 4K image isn't really more detailed than scaled to FHD but slower. Reduce brightness/gamma, modify RGB balance from yellowish tone back to original. On the flipside it shifts the colors but I'm happy to revert that back in FFmpeg. Restoration of black anime lines is great and the best I've seen. The clarity and image quality of that model are astounding. ![]() My favorite is "Anime Oldies" (the first one, not alternative) and is apparently based on 4xPSNR. Then within the ESRGAN family I've tried most of what's listed in the Manga / Anime subsection here: Super well designed by a great programmer, amazing handling of Python and other dependencies, free (MIT License), large selection of models Practically unusable for a full movie, slow and cumbersomeįast but hardly any better than simple bicubic upscale with post-processing, has a rather blurry oil painting lookĭecent results, well designed, comparably fast for what it achieves, doesn't artefact at allĭecent product, sharper than Waifu2x, works better on some scenes than others, artefacting can be controlled with pre-processing in Hybrid and selection of models in VEAI Here's what I've tried to upscale hand drawn 80's anime, in order from worst to best: ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |