Even top image generators lose track of who they're depicting.

Generative image tools create great pictures, but they struggle to keep the same person or object looking consistent. When an image shifts a subject's pose, viewpoint, expression, or scene, their identity often drifts. Researchers tested different ways systems store identity: in the prompt context, inside trained model settings, or as a persistent identity layer that gets reused. They tested these setups across generation, editing, fixing damaged images, and multi-subject scenes. They found that high image quality and good prompt following do not guarantee identity stays true. Identity degrades even faster during repeated edits, small scales, heavy damage, and multi-subject pictures. However, using a persistent identity layer substantially reduces this loss across generation, editing, and restoration. It works across different base models while keeping quality and prompt following high. Storing identity as separate, persistent knowledge preserves who or what is depicted.

Why it matters

For artists, designers, or anyone using AI to create images of specific people, current models often change the subject's identity with changes in pose or background. This makes it difficult to maintain consistency across multiple images or edits, which is crucial for storytelling and character design.

Questions answered