Back to feed
arXiv cs.AI·

IdGlow: Dynamic Identity Modulation for Multi-Subject Generation

Signal
72
Hype
28
In three linesIdGlow is a mask-free Flow Matching framework for multi-subject generation. It uses task-adaptive timestep scheduling and temporal gating to preserve identities during complex transformations (e.g., age transformation). A fine-grained group-level DPO stage eliminates artifacts and harmonizes textures.
Read source
Your take?
Image generationFine-tuningReinforcement learningPapers

Summary generated by Claude — human-verified