Course contentsShow
← Machine Learning and Deep Learning
Lesson 1600 of 3,53834. Diffusion ModelsFree lesson

Consistency Models

Self-consistent models that map any point on a diffusion trajectory directly to the origin.

Consistency Models

What you'll learn: How consistency models enable direct one-step generation from diffusion processes by enforcing self-consistency across the entire trajectory.

The Core Idea

Traditional diffusion models require many denoising steps to go from pure noise back to a clean image. Consistency models solve this by learning a special function that maps any point along the diffusion trajectory directly to the data origin (the clean sample).

Think of it like this: imagine you're hiking down a mountain trail (the diffusion trajectory). Normally, you'd take hundreds of small steps. A consistency model is like having a teleportation device that works from any point on the trail—whether you're at step 5, step 50, or step 500—and instantly takes you to the bottom. The key constraint? No matter where you start on that trail, you should arrive at the same destination.

Self-Consistency: The Magic Property

The "consistency" in consistency models refers to a mathematical property: for any two points on the same diffusion trajectory, the model must map them to the same final output. If you denoise an image from timestep t=800 or from timestep t=400, you should get the identical result.

This self-consistency constraint is enforced during training, creating a function that "shortcuts" the entire reverse process. Once trained, you can sample in just one step (or a few steps for higher quality) instead of hundreds.

Why This Matters

Consistency models bridge the gap between diffusion quality and fast generation. They eliminate the need for progressive distillation's teacher-student setup while maintaining the ability to trade-off speed and quality by optionally taking more steps.

Key Takeaway: Consistency models learn a self-consistent mapping that can jump from any point on a diffusion trajectory directly to the final clean sample, enabling fast one-step generation while preserving multi-step refinement as an option.