Shulou(Shulou.com)12/24 Report--

Thanks to CTOnews.com netizen OC_Formula for the clue delivery! CTOnews.com, November 30, Stability AI recently launched Stable Diffusion XL Turbo (SDXL Turbo), an improved version of the previous SDXL model, which claims to use "adversarial Diffusion Distillation (Adversarial Diffusion Distillation)", reducing the iterative step of image generation from the original 50 steps to 1 step. It is said that "a single iterative step can produce a high-quality image."

It is reported that the biggest feature of the Stable Diffusion XL Turbo model is the above-mentioned "image generated by an iteration", which claims to be able to carry out "instant text-to-image output" and ensure the quality of the picture.

Among them, the "antagonistic diffusion distillation technology" uses the existing large-scale image diffusion model as the "teacher network" to guide the model generation process, combining "distillation technology" and "confrontation training". Among them, "distillation technology", that is, "distillation model knowledge, refining model output", condenses the knowledge of large models into smaller models, while adversarial training can improve the model. In order to better imitate the teacher model output.

Previous model distillation techniques are difficult to strike a balance between efficiency and quality, because fast sampling usually weakens output quality, so this Stable Diffusion XL Turbo model is an important step forward in generating high-quality images efficiently through "adversarial diffusion distillation".

Officials compare Stable Diffusion XL Turbo with several different model variants, including StyleGAN-T++, OpenMUSE, IF-XL, SDXL and LCM-XL, and conduct two experiments. The first experiment requires model evaluators to randomly look at the output of the two models and select the output image that best matches the prompt, while the second experiment performs roughly the same as the first experiment. Model evaluation requires the selection of the one with the best image quality of the output in the model.

The experimental results of ▲ image source Stability AI blog show that Stable Diffusion XL Turbo can greatly reduce the computing requirements while still maintaining good image generation quality. The model word iteration is better than four iterations of LCM-XL, while four iterations of Stable Diffusion XL Turbo can beat the previous 50-step iterative configuration of Stable Diffusion XL; on the A100 GPU to calculate the 512x512 resolution image in 207ms.

CTOnews.com noted that Stability AI has published the relevant code on Hugging Face for personal and non-commercial use. Friends who are interested can click here.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.