Leco score v1: https://civitai.com/articles/4216

Leco score v1 sample: this

Leco score v2: https://civitai.com/articles/5416

Leco score v2 sample: https://civitai.com/models/471936

Leco score v2 training guide: https://civitai.com/articles/5422

A proof of concept Lora for LECO_score.

The Lora has been trained by generating images on AnyLora with prompts in PartiPrompts, asking the Lora to optimise the aesthetic score (in the sense of the AVA dataset, which is a dataset of photos evaluated by professional photographers).

The result is quite good with some caveats:

• more soft shadows

• more vibrant colours

• more interesting camera angles and perspective

• better colour harmony

• central subject of the photo pops out more from its background

• BUT it likes arty photos with abstract backgrounds, and will not interact well with detailed backgrounds

• BUT it dislikes backlighting (it's bad photography) so models that rely on backlighting and edge lighting like ComiNoir will get weird results

• BUT it was weird to train a score on realistic photos and then use it on anime, I don't care

Goods results with DreamShaper so far.

Training scripts, data, intermediate resnet checkpoints attached.