Text-to-image models like Stable Diffusion produce images step by step, making them slow and expensive to run. Distillation compresses them into a single step, but current methods are complex, requiring multiple training objectives and auxiliary models. A simpler approach, adversarial training, has struggled at modern scale due to the critic, the network that judges image realism, being poorly adapted to today's visual representations.
The authors of ROAD redesign the critic: they replace its simple convolutional head with a transformer head that can better interpret modern visual features, periodically refresh a random feature mixing to prevent the generator and critic from over-adapting to a single view, and introduce a new penalty (Jacobian R1) that stabilizes training for the critic's many output dimensions. All these changes let a single adversarial objective suffice.
On the COCO-10K benchmark, ROAD reached an FID of 11.03 with SDXL and 13.73 with SD3.5-M, scores that beat the multi-step teachers at a fraction of the training compute. This makes one-step distillation dramatically simpler and faster to train, bringing high-quality one-step generation closer to practical deployment.