This model was trained on a dataset of 20 images. The training went on for 10 epochs with 5 repeats each. There is no HD version of the cartoon, so I took screenshots of the DVD(720x576) and upscaled them using R-SERGAN 4x+.
General information:
Best weight: 0.7-0.9
Appearance: dolly, 1girl, flat chest, green eyes, orange hair, twintails
Outfit: pink dress, white socks, grey mary janes
Highly recommended to use LoRA for a different style.
Description
FAQ
Comments (1)
You don't actually need to upscale old cartoons, so long as you can get the minimum 250x250 resolution and the edges of the character (and their outfit parts) are clearly visible. Upscaling everything will cause the model to learn upscaling artifacts (flat colors, overly smooth and thick outlines). It makes more sense when you learn about how training processes images based on their "depth".
Instead try tagging images as "lowres", "blurry" and (optionally) "19?0s (style)" as applicable. It teaches it those are not part of the character but part of "lowres, blurry" (since it knows those concepts already and can compare it against that). This works best if you have some higher quality images (e.g., promo material, fan art) that doesn't get the tags. You can see it in action with my Atsuko model.