eval

frechet inception distance — measures distributional similarity between two sets of images

clipscore — cosine similarity between clip text and image embeddings for prompt adherence

arcface identity similarity — cosine similarity between face embeddings of two images

pickscore v1 — scores how well an image matches a text prompt (clip-h finetuned on pick-a-pic)

crop and resize images — center crop to square, face-aware crop, or custom resize

degrade images with blur, noise, and compression for benchmark testing

research-quality chart generation (bar, grouped bar, scatter) using matplotlib