AI-powered image generation from a simple text prompt has reached new speeds, thanks to the groundbreaking methods developed by Stability AI. The introduction of their Stable Diffusion model has revolutionized the process, eliminating the need for humans to wait precious seconds or minutes for image generation.
Introducing SDXL Turbo: Real-Time Image Generation for the Masses
Stability AI has recently unveiled SDXL Turbo, a game-changing feature that enables real-time image generation. By significantly reducing the number of generation steps required, users can now experience instantaneous results. Previously, 50 steps were needed, but now only one is necessary. This major development also reduces the compute load, making the process more efficient and accessible to all.
“SDXL Turbo can generate a 512×512 image in just 207ms on an A100 GPU, which is a major speed improvement over prior AI diffusion models.” – Stability AI
The speed at which SDXL Turbo operates is reminiscent of the predictive typing capabilities now found in popular search engines, browsers, and operating systems. However, SDXL Turbo takes this concept to a whole new level, allowing users to generate images at the speed of thought.
“One step Stable Diffusion XL with our new Adversarial Distilled Diffusion (ADD) approach. Less diversity, but way faster & more variants to come which will be… interesting, particularly with upscales & more..” – Emad Mostaque, founder and CEO of Stability AI
The Technology Behind SDXL Turbo
SDXL Turbo is powered by a revolutionary technique called Adversarial Diffusion Distillation (ADD). Unlike relying on superior hardware, Stability AI’s ADD approach utilizes a new method for accelerating SDXL. Through combining single-step Stable Diffusion XL with Adversarial Distilled Diffusion, SDXL Turbo achieves unmatched speed without compromising quality or accuracy.
The original SDXL base model laid the foundation for this breakthrough development. With ControlNets, a concept introduced by Stability AI, better control over image composition is achieved. Furthermore, the incorporation of 3.5 billion parameters enhances the model’s accuracy by increasing its awareness of various concepts.
Redefining Generative AI: Development Pathways
Stability AI follows a common path in modern generative AI development. Starting with the creation of the most accurate model possible, they then optimize it for performance. This trajectory is similar to OpenAI’s approach with GPT 3.5 Turbo and GPT-4 Turbo. While there is typically a tradeoff between speed and quality in accelerated generative AI models, SDXL Turbo minimizes this compromise. The results produced are highly detailed, with only a marginal decrease in image quality compared to the non-accelerated version of SDXL.
The ADD research report highlights the significance of marrying the best aspects of diffusion models and GANs. Stability AI’s ADD approach is designed to outperform other AI methods for image generation, leveraging adversarial training and score distillation.
“The aim of this work is to combine the superior sample quality of diffusion models with the inherent speed of GANs.” – ADD Research Report
The groundbreaking Adversarial Diffusion Distillation (ADD) technique developed by Stability AI researchers harnesses the power of pretrained image diffusion models. By using adversarial training and score distillation, ADD enables fast sampling, iterative refinement, and the utilization of stable diffusion model pretraining. In comparative experiments, ADD outperformed GANs, Latent Consistency Models, and other diffusion distillation methods in 1-4 steps.
Preview and Future Possibilities
While SDXL Turbo is not yet ready for commercial use, Stability AI has made it available for preview on their Clipdrop web service. Early tests conducted by VentureBeat reveal the impressive speed of image generation, although certain advanced parameter options for different styles are currently unavailable in the Clipdrop beta version. Stability AI has also released the code and model weights for SDXL Turbo on Hugging Face, under a non-commercial research license.
