What they did

They designed a benchmark covering subject-driven generation, editing, restoration, and multi-subject settings, with tasks that increasingly stress identity preservation. They compared three paradigms: input-context methods (e.g., GPT-Image-2, NB2), subject-specific fine-tuning (LoRA), and a persistent identity layer (Phota Identity). The evaluation measured identity fidelity, image quality, and instruction adherence.

Key findings

• Identity preservation is a distinct limitation: strong image quality and instruction following do not guarantee high identity fidelity.
• Identity degradation worsens under iterative edits, small subject scales, severe image degradation, and multi-subject composition.
• Persistent identity layers substantially reduce identity degradation across all tasks and consistently improve identity preservation when applied to different foundation models.
• Instruction adherence and perceptual image quality remain comparable to baseline methods.

Why it matters

This work demonstrates that identity is not automatically captured by increasingly capable generative models; instead, it can be represented as persistent subject knowledge that is composed independently with the generative model. This insight could guide the design of future generative systems that reliably preserve identity across diverse contexts.

Caveats

The benchmark focuses on a set of existing methods; the authors note that the results may depend on the specific implementations of each paradigm. The persistent identity approach is evaluated only on a single proprietary system (Phota Identity), so generalizability to other persistent identity representations is not established. Real-world applications may involve additional challenges not covered in the benchmark, such as identity preservation in video or across very long sequences of edits.