SDXL Turbo is an innovative text-to-image model that leverages Adversarial Diffusion Distillation (ADD) to produce images in a single step, enhancing efficiency by reducing the traditional 50-step process to one.
The model’s distillation approach merges adversarial training with score distillation, as Stability AI’s research paper elaborates, allowing for high-quality image generation free from artifacts and blurriness.
In comparative tests, SDXL Turbo outshines other diffusion models like StyleGAN-T++, OpenMUSE, IF-XL, SDXL, and LCM-XL in prompt adherence and image quality, even besting LCM-XL’s 4-step and SDXL’s 50-step processes with fewer steps. It showcases exceptional image quality with reduced computational demands.
Furthermore, SDXL Turbo achieves remarkable inference speeds, generating 512×512 images in just 207ms on an A100 GPU, and is compatible with Stability AI’s Clipdrop platform, inviting users to experience its real-time generation capabilities. However, SDXL Turbo is not yet cleared for commercial use, and interested parties should reach out to Stability AI for more details.
