This work three paradigms for preserving subject identity in generative image models: in-context identity (e.g., GPT-Image-2), model-specific identity via LoRA, and persistent identity via a dedicated identity layer. The evaluation spans subject-driven generation, editing, restoration, and multi-subject scenarios, with tasks that increasingly stress identity preservation.
The results show that strong image quality and instruction following do not guarantee identity fidelity. Identity degradation worsens under iterative edits, small subject scales, severe image degradation, and multi-subject composition. Persistent identity substantially reduces this degradation across all tasks while maintaining comparable instruction adherence and perceptual image quality.
These findings challenge the assumption that identity naturally emerges from more capable generative models. Instead, identity can be represented as persistent subject knowledge that composes independently with the underlying generative model, pointing toward decoupled identity modules for reliable subject consistency.