task-amenability VS improved-aesthetic-predictor

The repo: https://github.com/s-sd/task-amenability

As a sidenote, we have been working on an approach for data valuation which does not require a clean validation set. You can access the paper and code and try that out if you'd like.

Perhaps you could use an "aesthetic-evaluator": https://github.com/christophschuhmann/improved-aesthetic-predictor (probably not SoTA but it works)

"Calculates aesthetic score for generated images using CLIP+MLP Aesthetic Score Predictor based on Chad Scorer "

stable-diffusion-v1-2: Resumed from stable-diffusion-v1-1. 515,000 steps at resolution 512x512 on "laion-improved-aesthetics" (a subset of laion2B-en, filtered to images with an original size >= 512x512, estimated aesthetics score > 5.0, and an estimated watermark probability < 0.5. The watermark estimate is from the LAION-5B metadata, the aesthetics score is estimated using an improved aesthetics estimator).

The data used for fine-tuning has come from a random sample of 56k Danbooru images, which were filtered based on CLIP Aesthetic Scoring where only images with an aesthetic score greater than 6.0 were used.

All they've said is they randomly picked 56,000 images that had an aesthetic score greater than 6.0. The score is created by this model. https://github.com/christophschuhmann/improved-aesthetic-predictor

Does anyone know how you could combine a model like Aesthetic Score Predictor with stable diffusion? They used this model to filter training images by score. Iit seems like a lot of people just tune their prompt to make images more aesthetic, by adding certain words.