Identity-Preserving Image Generation
Overview
Researchers condition diffusion and flow-based models on face-recognition embeddings, reference-image features, or personalized adapters, often adding attention, spatial masks, and identity losses to preserve faces and broader appearance while following prompts. Other approaches fine-tune models, optimize outputs at test time, or use reward-based training to balance identity with requested changes. Evaluation combines recognition similarity and perceptual quality with prompt alignment and human judgments, increasingly testing edits, multiple people, and attribute binding. The main tension is that stronger identity control can limit diversity or prompt flexibility and cause copy-paste artifacts, while performance can vary across demographic groups and larger scenes.
Latest papers
48 papers3 papers in the last four weeks, against 2 the four weeks before. 0.0% of all new papers.