Stability AI launches Stable Diffusion XL Turbo model

IT House reported on November 30 that Stability AI recently launched Stable Diffusion XL Turbo (SDXL Turbo), which is an improved version of the previous SDXL model. It is said that SDXL Turbo uses "Adversarial Diffusion Distillation technology" to reduce the iteration steps of image generation from the original 50 steps to 1 step. It is said that "only one iteration step is needed to generate high-quality images." Image”

It is reported that the biggest feature of the Stable Diffusion XL Turbo model is the above-mentioned "generating images in one iteration", which is claimed to be able to perform "instant text-to-image output" and ensure the quality of the images.

What needs to be rewritten is: One of them is called "adversarial diffusion distillation technology", which is a technology that uses the existing large-scale image diffusion model as a "teacher network" to guide the generation process. This technology combines "distillation technology" and "adversarial training", where "distillation technology" refers to condensing the knowledge of a large model into a smaller model to streamline the output of the model. Adversarial training can improve the model so that it can better imitate the output of the teacher model

Previous model distillation technology was difficult to balance efficiency and quality because fast sampling usually weakens the output quality. Therefore, this Stable Diffusion XL Turbo model uses "adversarial diffusion distillation technology" to efficiently generate high-quality images. It is an important progress.

Officially compared Stable Diffusion XL Turbo with multiple different model variants, including StyleGAN-T, OpenMUSE, IF-XL, SDXL and LCM-XL, and conducted two experiments. The first experiment required the model The evaluators randomly view the output of the two models and select the output image that best matches the prompt word. The second experiment is performed roughly the same as the first experiment. Model evaluation requires selecting the one with the best image quality in the model. .

▲ Picture from Stability AI Blog

Experimental results show that Stable Diffusion XL Turbo can significantly reduce computing requirements while still maintaining excellent image generation quality. This model surpasses LCM-XL in just one iteration compared to 4 iterations. The Stable Diffusion XL Turbo, which has gone through 4 iterations, can easily beat the Stable Diffusion XL that previously required 50 iterations to configure. Computing an image at 512x512 resolution in just 207 milliseconds using the A100 GPU

IT House noticed that Stability AI has currently published the relevant code on Hugging Face for personal and non-commercial use. Interested friends can click here to visit.

The above is the detailed content of Stability AI launches Stable Diffusion XL Turbo model. For more information, please follow other related articles on the PHP Chinese website!