AniToon is a SISR model trained to faithfully restore low-quality 90s or 2000s cartoons and anime (much like LD Anime). The dataset is a bit of an abomination, with over 50 different sources cobbled together into a Frankenstein's monster of a dataset. The result is a model with excellent restoration capabilities on older/low-quality inputs, while still maintaining faithfulness to the source.
The model comes in three sizes: RPLKSR Small, RPLKSR and RPLKSR Large. Small should be sufficient for most use cases, but if you need a bit more punch, the normal full sized model and the large versions can be useful.
- Comparisons: Slow.pics
- Training Configs: The configs used to train the models can be found here (neosr).
Warning
One major caveat to all this is that there appear to be compatibility issues between the model and VSMLRT. I've attempted to debug the problem, to no avail. To elaborate, the TensorRT and DML output in VSMLRT materially deviates from the correct/intended Pytorch output. As a result, the model is only shared in PTH format, for inference in Chainner (recommended for single images) or Real Video Enhancer (recommended for videos, using Pytorch backend).
Thanks to kuronoe, sharekhan and pwnsweet for their significant contributions to the dataset! 💖