vicgalle/stable-diffusion-aesthetic-gradients

Implements gradient-based optimization during the diffusion process to steer generation toward user-defined aesthetic embeddings derived from CLIP, eliminating the need for prompt engineering. Computes aesthetic gradients at each denoising step by optimizing a frozen CLIP embedding representing desired visual characteristics, with tunable `aesthetic_steps` and learning rate parameters. Includes pre-computed embeddings from curated datasets (LAION, SAC, artist styles) and provides `gen_aesthetic_embedding.py` to generate custom embeddings from arbitrary image collections.

741 stars. No commits in the last 6 months.