NearID: Identity Representation Learning via Near-identity Distractors
The research introduces Near-identity (NearID) distractors—semantically similar but distinct images placed on the same background—to test how well AI models capture true identity versus background context. By training a lightweight adapter with a hierarchical contrastive objective, the method successfully isolates intrinsic identity signals. This results in significantly improved performance in tasks like image editing and personalization by making representations more robust to contextual shortcuts.