11 upvotes
VA-TDM boosts 4-step diffusion image quality with adversarial distillation
What is VA-TDM: A Unified View for Improving Few-Step Diffusion Models by Velocity Adversarial Distillation?
The authors propose Velocity Adversarial Distillation, a way to train a fast diffusion sampler by matching its velocity field to that of a full-step model while using an adversarial loss to tighten the distribution match. The result is high-fidelity images generated in only a few sampling steps.
Who it is for
Anyone building real-time image generators or low-latency AI services can use VA-TDM to slash inference time without sacrificing visual quality.