✦
Summary
DMAD is an adversarial distillation framework for reducing the sampling steps of visual generators. Two discriminator heads on a shared backbone classify teacher, student, and real samples, using the resulting distribution gaps to train a few-step student.
Key Features
Visual-model distillation.
Two discriminator heads.
A shared discriminator backbone.
Teacher-student matching.
Real-data supervision.
One-step image experiments.
Few-step video generation.
Released code and models.
About DMAD
Generative-model researchers can adapt the released framework to image or video teachers while studying quality versus inference cost. Core capabilities include Visual-model distillation, Two discriminator heads, A shared discriminator backbone, and Teacher-student matching.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Real-data supervision, One-step image experiments, Few-step video generation, and Released code and models.