Open Enferlain opened 4 months ago
pick scorer_v1 + aesthetic_shadow_v2 scorers added.
@wkpark You added the original shadow scorer which gives problematic scores, hence the v2.
I tested multiple images, both viable and messed up ones, including some that were 80% noise. Every single image was given a 0.99+ score.
@Enferlain I was looking for scorers myself so I just made a guide on how to do it.
There are a few potential scorers that could be worth a try
https://huggingface.co/shadowlilac/aesthetic-shadow-v2 - Anime. Large model, maybe needs to be converted to fp16. Was used for the Animagine v3 datasets https://huggingface.co/Eugeoter/waifu-scorer-v2 - Anime. Small. By one of the better anime finetune makers. Haven't tested yet. https://huggingface.co/yuvalkirstain/PickScore_v1 - Large model, not 100% sure about the function but I've used it before.
Idea from here along with some other implementations like BLIP CLIP hpsv2 etc.
Maybe could run multiple scorers together and average them out to have more opinions. And it could also be nice to implement a manual scoring option, although that wouldn't strictly fall under "auto merging" I guess, but it would be useful nonetheless.