Adversarial Diffusion Distillation

Key Insight

Pure few-step distillation tends to produce blurry images, because regressing toward an average washes out fine detail. Adversarial Diffusion Distillation (ADD) — the recipe behind SDXL Turbo — fixes this by adding a GAN-style discriminator that rejects any quick output which doesn't look real, forcing the 1–4-step student to stay sharp. Comparing it head-to-head with an LCM shows the trade-off plainly: ADD's adversarial training is fiddlier to run but yields crisper few-step samples.