Adversarial Diffusion Distillation and Turbo Models

A standard diffusion model creates an image by removing noise over many steps, commonly 20 to 50, and each step is a full pass through a large network. Distillation trains a student network to reach a similar result in far fewer steps. Stability AI introduced Adversarial Diffusion Distillation in a November 2023 paper by Axel Sauer and colleagues, released alongside SDXL Turbo. The student starts from the pretrained SDXL weights and is trained with two losses. The adversarial loss comes from a discriminator that tries to tell the student's outputs from real images. In ADD, this discriminator is built on a frozen pretrained vision backbone (DINOv2) with small trainable heads. The distillation loss uses the original SDXL as a teacher. The student's output is noised again, the teacher denoises it, and the student is pushed toward the teacher's prediction, a form of score distillation. The adversarial term keeps single-step images sharp, and the teacher term keeps them faithful to what the large model knows. A later variant, Latent Adversarial Diffusion Distillation (LADD), runs the discriminator in latent space using the teacher's own features, which avoids expensive decoding to pixels. It was used for SD3 Turbo, and Black Forest Labs has said Flux.1 schnell was trained this way. The trade-offs are real. Distilled models usually lose some sample diversity, so different seeds for the same prompt can look alike. Classifier-free guidance is typically built in, so the guidance scale and negative prompts behave differently or barely work. Fine detail, text rendering and unusual compositions may lag behind the teacher. A common misconception is that a turbo model is just the full model run with fewer steps. A normal model run at one step gives a blurry mess, and the speed comes from retraining. Other routes to fast sampling include consistency models, LCM, progressive distillation and SDXL Lightning.