metadata
license: mit
base_model:
- ostris/vae-kl-f8-d16
Another take on upscaler using AsymmetricAutoencoderKL, in this case I did re-use decoder of ostris/vae-kl-f8-d16 as decoder and trained just encoder.
Oversharpening was reduced by usage of sobel in loss calculation. I guess that without high resolution image used as reference, it might look more blury than previous one.