[09:05 Wed,6.December 2023 by Rudi Schmidts] |
When reporting on current advances in generative AI models, it is sometimes difficult not to use the term game changer. In the case of the current development progress of Stable Diffusion, it is even extremely difficult, because what the team from Stabilty AI has recently presented has yet to be discovered by the assembled competition.
When working with  ComfyUI - currently the hottest interface for stable diffusion - images can be output every second with fast GPUs. In other words, practically in real time, while you are still typing your prompt. So you can immediately see what changes even one letter more or less in the prompt causes: If you want to try it out for free, you can go directly to use an online prompt here at Stability AI. But be careful, use with caution - the free trials are already "used up" after approx. 30 actions. If you are interested in exactly how the new algorithm works, you can find here is the corresponding paper on "Adversarial Diffusion Distillation". The new ADD idea seems to borrow from Generative Adversarial Networks (GANs). ;) SDXL-Turbo - New AI algorithm extremely accelerates stable diffusion If you want to install and run SDXL Turbo yourself, we recommend the following speed data points: With an Nvidia RTX 3060, you can expect images with a resolution of 1,024 x 1,024 pixels to be ready in around 4 seconds after three steps. On an Nvidia A100, a 512 x 512 pixel image only takes around 200 milliseconds according to Stability AI. And here you can see on Reddit how SDXL Turbo runs on an RTX 4090.  more infos at bei comfyanonymous.github.iodeutsche Version dieser Seite: SDXL-Turbo - Neuer KI-Algorithmus beschleunigt Stable Diffusion extrem |
 |
|
|