The previous release was meant to be as trouble-free as possible, and to serve as a good starting point for subsequent improvements. This marks the first minor release, bringing reconstruction quality to roughly where it was a few weeks ago before I started tweaking the models, but without any of the reported issues!
Changes:
- Training dataset now contains both anime frames and fanart at a 75%/25% split. This helps reconstruction quality on very detailed frames without introducing weird artifacts on meshes or credits.
- All models are now trained using AdamW instead of Adam. This seems to help a little bit with generalisation, but it could be placebo.