SDXL Turbo

SDXL Turbo is a new text-to-image generation model that utilizes a distillation technique called Adversarial Diffusion Distillation (ADD), allowing it to create image outputs in a single step. This real-time model maintains high sampling fidelity while significantly reducing the required step count from 50 to just one, resulting in improved efficiency. The distillation technique used in SDXL Turbo combines adversarial training and score distillation, as detailed in the research paper provided by Stability AI.

This technique enables the model to generate image outputs with high quality and eliminates common issues such as artifacts and blurriness that may occur with other distillation methods. In performance comparisons with other diffusion models like StyleGAN-T++, OpenMUSE, IF-XL, SDXL, and LCM-XL, the tool outperformed these models in terms of following the given prompt and image quality, even surpassing a 4-step configuration of LCM-XL with just a single step and a 50-step configuration of SDXL with only 4 steps. This demonstrates the tool's ability to provide superior image quality while reducing computational requirements.

Additionally, the tool offers enhanced inference speed, allowing it to generate a 512x512 image in just 207ms on an A100 GPU. Its compatibility with Stability AI's image editing platform, Clipdrop, provides users with an opportunity to explore and test the capabilities of this real-time image generation model. Please note that the tool is currently not intended for commercial use, and if you're interested in using this model for commercial purposes, you should contact Stability AI for further information.