Get Started
Home
Topics
Search
Library
Research questionHow can reinforcement learning post-training for diffusion models avoid objective mismatch and high-variance updates?Diffusion models are commonly pretrained with score- or flow-matching objectives, while some reinforcement-learning post-training methods optimize a different objective. This mismatch can produce noisy estimators, increase variance, and slow convergence.
Diffusion Models
Image Generation
Machine Learning
Reinforcement Learning
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.Advantage Weighted Matching: Aligning RL with Pretraining in Diffusion ModelsApplies to reinforcement-learning post-training of diffusion image generators. The evidence includes theoretical analysis and results on GenEval, OCR, and PickScore using Stable Diffusion 3.5 Medium and FLUX, including comparisons with Flow-GRPO.research paper · Sep 1, 2026
Related questions
How can diffusion image and video generators be preference-aligned without inefficient training exploration or inference-time search?How can on-policy diffusion RL reuse rollout computations without trading redundant recomputation for excessive memory use?How can text-to-image diffusion models use multiple candidates and continuous rewards beyond pairwise preferences?How can conditioned diffusion models be guided without score functions when diffusion coefficients are singular and conditional densities are absent or nonsmooth?