DiffusionNFT

DiffusionNFT: Online Diffusion Reinforcement with Forward Process

Kaiwen Zheng et al. · ICLR 2026 Oral

Read paper · arXiv:2509.16117

Construct positive and negative implicit velocity fields from feedback to optimize diffusion models.

What you will implement

Positive/negative beta implicit velocities, the two-branch objective, and tiny training.

• 49. DiffusionNFT Positive and Negative Flow Matching

  Construct implicit positive and negative policies from one trainable velocity and optimize both with probabilistic rewards.

  Implement the Eq. (5) training core; forward noising and tiny training scaffolding are supplied. Image reward models, online rollouts and EMA updates are excluded.