DMAD: Distribution Matching as Adversarial Distillation for Fast Visual Generation
Presents breakthrough technique for accelerating diffusion models with theoretical guarantees and empirical results
AI Summary
Researchers introduce DMAD, a novel adversarial distillation method that achieves state-of-the-art few-step visual generation performance on ImageNet-64x64, SDXL, and video generation benchmarks.
Excerpt
Distribution Matching Distillation (DMD) trains a few-step student from the difference between separately estimated target and student scores, so it must keep an auxiliary diffusion model fitted to the student's evolving distribution at extra memory and computation cost. We introduce DMAD, Distribution Matching as Adversarial Distillation, which recasts distribution matching as classification and learns the required log-density ratios directly. Two discriminator heads on a shared backbone distin