SDXL Turbo (Adversarial Diffusion Distillation)
Adversarial Diffusion Distillation makes SDXL usable in 1-4 steps, claimed as the first real-time single-step synthesis from a foundation model.
A student model is trained with score distillation from a large teacher plus an adversarial discriminator loss. The authors report it beats GANs and Latent Consistency Models at one step and matches SDXL quality within four steps, opening interactive generation.
- Date
- Tuesday, 28 November 2023
- Lab
- Stability AI
- Kind
- paper
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Sampling steps | 1-4 leading SDXL-level quality within 4 steps per abstract | authors |
Date is the arXiv submission (2023-11-28); Wikipedia places the SDXL Turbo weights release in the same month. Access terms of the weights not re-checked.
Sources
This record was checked against its sources on 6 October 2026. How we check