Skip to content

SDXS

A real-time one-step latent diffusion model

Visit official site

What SDXS is

The model is described as a diffusion model that lowers latency through model miniaturisation and fewer sampling steps, using knowledge distillation to simplify the U-Net and the image decoder.

What you can do with it

  • Generate images at high speed
  • Control generation with image conditions
  • Translate one image into another
  • Colourise and reconstruct sketches or old photos

Model notes

  • The SDXS-512 variant reaches about 100 FPS on a single GPU
  • The SDXS-1024 variant reaches about 30 FPS on a single GPU
  • Both figures are stated as 30 to 60 times faster than earlier models
  • Single-step training uses feature matching and score distillation

Application scenarios

  • Generate a landscape image with SDXS-512
  • Colourise and reconstruct a sketch image
  • Automatically colourise an old black-and-white photo

Pros & cons

✓ What we like

  • The material states about 100 FPS for SDXS-512 and 30 FPS for SDXS-1024 on a single GPU
  • Image-conditioned generation and image-to-image translation are both described

! What to watch out for

  • The stated frame rates are tied to a single GPU setup
  • Image-to-image translation is described as a potential application of the training method

FAQ

How fast are the two models?

The material states about 100 FPS for SDXS-512 and 30 FPS for SDXS-1024 on a single GPU.

What techniques are named?

Knowledge distillation, single-step training with feature matching and score distillation, and image-conditioned control.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review