Making slow image generators 100× faster with one-step distillation

Discrete diffusion models generate high-quality images but require slow iterative decoding. Fixed-Point Distillation (FPD) compresses this into one inference step by training a student model to mimic a teacher's corrections. The key trick: corrupt the student's draft slightly, let the teacher refine it once, then train the student to predict that refined version—all while routing gradients through discrete tokens via a straight-through estimator. On class-conditional and text-to-image tasks, FPD matches multi-step teachers while being orders of magnitude faster at inference.