Consistency-Model Distillation

Key Insight

A diffusion model needs around 50 denoising steps to produce a clip, which is far too slow for production use; distillation trains a fast "student" to reproduce that 50-step "teacher" in only about 4 steps. A consistency model is the student design that makes this possible: it is trained so that every point along the noisy-to-clean denoising path points straight at the same finished clip, so one or a few jumps replace the long walk. This project performs that distillation and measures the bargain directly — the speed you gain (50 ÷ 4 ≈ 12× fewer steps) against the quality you lose. Where the real-time latency hunt project chases raw milliseconds-per-frame for interactivity, this one zeroes in on the speed-versus-quality trade of the distillation step itself.