Skip to content

DiffusionRL

Apply reinforcement learning to diffusion models

Visit official site

What DiffusionRL is

The paper is described as a scalable reinforcement learning approach that improves text-to-image diffusion models with diverse reward functions.

What you can do with it

  • Apply reward functions for preference, composition and fairness
  • Align a diffusion model with human preference
  • Improve composition and diversity of generated samples

Pros & cons

✓ What we like

  • The method is described as scalable across millions of images
  • Composition and diversity of samples are reported as improved

! What to watch out for

  • The work targets biases from web-scale text-image training pairs
  • The named base model is a pretrained stable diffusion model

FAQ

What problem is described?

Hidden biases in web-scale text-image training pairs can lead to suboptimal samples.

What result is reported?

The paper reports samples preferred by humans at 80.3 percent on a pretrained stable diffusion model.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review