SDXL Turbo — real-time generation

SDXL Turbo uses adversarial diffusion distillation to collapse generation to a single step. On a decent GPU, it renders 512x512 images at 10+ FPS — fast enough that users see the image update as they type.

• Live prompt sandbox with instant preview
• Drawing tools where the model reimagines your strokes as you draw
• Classroom demos that show how prompts change output in real time
• Mood-board generation in Figma-like speed

Design tradeoffs

• Turbo trades top-end quality for speed — do not expect print-ready output
• Text and hands still struggle
• Works best at 512 or 768 resolution; higher sizes lose the speed advantage
• Pair with a second-pass refiner for finals

from diffusers import AutoPipelineForText2Image
import torch

pipe = AutoPipelineForText2Image.from_pretrained(
    "stabilityai/sdxl-turbo", torch_dtype=torch.float16
).to("cuda")

img = pipe(prompt="a fox in a meadow", num_inference_steps=1, guidance_scale=0.0).images[0]

Show Turbo live while the user iterates. When they commit to a prompt, render the final with SDXL base or Flux Pro for print-quality output. Best of both worlds — real-time feel, publish-grade finals.