Skip to content

Score Distillation Sampling

Use a diffusion model to guide text-driven optimisation

Visit official site

What Score Distillation Sampling is

The method is described as a way to use image diffusion models to steer optimisation problems guided by text prompts. The associated paper examines the loss and proposes a fix for the noisy gradient component.

What you can do with it

  • Run optimisation-based image synthesis and editing
  • Train a zero-shot image translation network
  • Apply the approach to text-to-3D synthesis

Application scenarios

  • SDS is used for image synthesis and editing
  • SDS is used to train a translation network
  • SDS is used for text-to-3D synthesis

Pros & cons

✓ What we like

  • The analysis identifies an issue in the original formulation
  • Qualitative and quantitative experiments are reported

! What to watch out for

  • High text guidance in the original formulation is linked to side effects
  • The fix relies on a trained shallow network

FAQ

What does the paper analyse?

The material states it analyses the SDS loss function.

What applications are listed?

The feature list names image synthesis and editing, zero-shot translation training and text-to-3D synthesis.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review