Stability AI has released their latest innovation in text-to-image generation technology - SDXL Turbo. The model utilizes a new technique called Adversarial Diffusion Distillation (ADD), which gives the model many advantages similar to Generative Adversarial Networks (GANs), such as single-step output generation while avoiding common artifacts or blurriness issues found in other distillation methods. In simple terms, SDXL Turbo can create detailed and high-fidelity images in real-time from short text prompts.

As detailed in Stability AI's research paper, the ADD technique allows SDXL Turbo to shorten the text-to-image process to just one step, whereas previous models like SDXL 1.0 required 50 steps to output an image. This significant efficiency improvement greatly reduces the computational power and time required for image generation without compromising visual quality.

In comparison tests with other state-of-the-art diffusion models, human evaluators consistently found that SDXL Turbo produces higher-quality image outputs with significantly fewer inference steps required compared to other models. By incorporating ADD technology, SDXL Turbo combines the strengths of diffusion models and GANs, eliminating common issues such as image blurriness or excessive smoothness.

In fact, the speed of SDXL Turbo is remarkable. On an A100 GPU, the model can generate a 512x512 pixel image in just over 200 milliseconds, including prompt encoding, denoising, and decoding. ADD distillation unlocks image generation speeds several orders of magnitude faster than multi-step methods, opening up new possibilities for real-time image creation using natural language prompts.

However, the currently released version of SDXL Turbo does have some notable limitations. The image resolution is fixed at 512x512 pixels, and the model may struggle to render clear text. Faces and human figures may not always be generated correctly. Therefore, while SDXL Turbo has made rapid progress in text-to-image generation with its innovative ADD technology, appropriate expectations need to be adjusted.

If you are interested in trying the SDXL Turbo model, Stability AI has made it available for free on Clipdrop. The beta version demonstration showcases the model's real-time text-to-image generation capabilities, accessible to most internet users.

SDXL Turbo is currently released under a non-commercial research license, allowing for non-commercial use by individuals. The model's weights and code are available on the Hugging Face platform.