Conventional reward fine-tuning often uses a reverse-KL penalty, implemented as a squared difference between the fine-tuned and pretrained models’ velocity fields.
Direct distribution rewards preserve diversity during diffusion model fine-tuning
The authors regularize generated samples themselves rather than only constraining the dynamics that produce them.
Top University
Romeo Passaro · Jeongjae Lee · Jong Chul Ye · Michael M. Bronstein · Yoshua Bengio · Alexander Tong · +1 more
University of Oxford · AITHYRA · KAIST · Mila – Québec AI Institute · Université de Montréal
Research Digest··2 min read
Passaro et al.
§