High-Speed, High-Resolution Image Synthesis Using Latent Adversarial Diffusion Distillation

Diffusion models are a core technology for image and video generation, but they have the drawback of slow inference speed. Recently, the Adversarial Diffusion Distillation (ADD) method was introduced with the goal of single-step inference, but it had issues with difficult and costly optimization due to its reliance on a fixed DINOv2 discriminator.

The new method, Latent Adversarial Diffusion Distillation (LADD), overcomes these limitations. Unlike pixel-based ADD, LADD leverages the generative features of a pretrained Latent Diffusion Model. This approach simplifies the training process while boosting performance, enabling high-resolution and various aspect-ratio image synthesis.

The research team applied LADD to Stable Diffusion 3 (8B) to implement SD3-Turbo. SD3-Turbo achieves performance on par with state-of-the-art (SOTA) text-to-image generation models with just 4 steps of unguided sampling.

Additionally, the scaling behavior of LADD was systematically analyzed, and its effectiveness was demonstrated across various applications such as image editing and inpainting.