Researchers introduce CLPIPS, a customizable image similarity metric that better aligns with human preferences in text-to-image generation workflows.

CLPIPS extends the existing LPIPS metric by fine-tuning it to match human judgments rather than relying solely on objective measures

Current metrics like LPIPS and CLIP often fail to capture user preferences in context-specific image generation tasks

The approach uses lightweight human-augmented fine-tuning to improve perceptual alignment with minimal computational overhead

CLPIPS positions similarity metrics as adaptive tools for human-in-the-loop text-to-image workflows, helping users iteratively refine prompts