IDeaL generates maximally informative samples via Dead Leaves optimization

mbsariyildiz · x · 2026-09-01

Researchers proposed IDeaL to address the need for teacher training data in multi-teacher distillation (e.g., UNIC), especially when data is private, licensed, or lost. The core idea treats pixels as learnable parameters, optimizing Dead Leaves-generated synthetic images through frozen teacher models (like DINO, iBOT). The resulting samples look unrealistic but are maximally informative for distillation.

Related event: ECCV Paper IDeaL Distills Four Vision Teachers with Zero Real Images(4 posts)→

Original post →

More from Models

Models channel →