Yesterday, I listened to my wife's husband Bob (the adult world is just so complicated) talk about the updates on the latest product. The SDXL Turbo released yesterday is pretty interesting. By the way, let me do a hard sell for Bob - he offers some AI-generated content (AIGC) painting courses. If you're interested, you can reach out to him. (I'm not getting paid for this advertisement; it's because I always get his fresh shares and research for free.) His contact info is Bilibili account: https://space.bilibili.com/498518999.
Back to SDXL Turbo, it is a real-time text-to-image generation model. Let’s first take a look at its performance:
https://stability.ai/news/stability-ai-sdxl-turbo
SDXL Turbo adopts the new Adversarial Diffusion Distillation (ADD) technology, achieving state-of-the-art performance. It brings single-step image generation quality to an unprecedented level, reducing the required number of steps from 50 down to just 1 step.
You can try it out on Stability's image editing website: https://clipdrop.co/stable-diffusion-turbo
I tried it myself, and the speed is very smooth. You can check out some comparisons provided in the paper:
This model can synthesize image outputs in a single step, generating real-time text-to-image outputs while maintaining high sampling fidelity. The paper link for the model is: https://stability.ai/research/adversarial-diffusion-distillation