Continuous Function

Legacy Concept Lab

Consistency Models: One-Step Diffusion

Generates images in 1-2 steps instead of 50-1000 steps

Concept 78 of 100Generative ModelsPhase 4
Open DemoAll Foundations
#78ConsistencyGenerative Models
key equationf(x_t, t) = x_0 \quad \forall t
Phase 4: Generative modeling familiesConcept 78 of 100

Why It Matters for Modern Models

  • Generates images in 1-2 steps instead of 50-1000 steps
  • Bridges the speed gap between diffusion quality and GAN speed
  • The "distillation" approach: learn to jump directly to the answer

What Tutorials Skip

What is still poorly explained in textbooks and papers:

  • Diffusion models trace a path from noise to image; consistency models learn to skip
  • Self-consistency = "any point on the trajectory should predict the same endpoint"
  • Trade-off: fewer steps = faster but lower quality; find the sweet spot

Interactive Visualization

Core Math (Optional Deep Dive)

If you want intuition first, start with the key equation and the visualization. Come back here for the full walkthrough.

Key Equation
f(xt,t)=x0∀tf(x_t, t) = x_0 \quad \forall t

Consistency function maps any point on ODE trajectory to origin:

f(xt,t)=x0∀t∈[0,T]f(x_t, t) = x_0 \quad \forall t \in [0, T]

Self-consistency property:

f(xt,t)=f(xt′,t′)for all t,t′ on same trajectoryf(x_t, t) = f(x_{t'}, t') \quad \text{for all } t, t' \text{ on same trajectory}

Training via consistency loss:

L=E[d(f(xt+Δ,t+Δ),f(xt,t))]\mathcal{L} = \mathbb{E}\left[d(f(x_{t+\Delta}, t+\Delta), f(x_t, t))\right]

One-step generation: x0=f(xT,T)x_0 = f(x_T, T) where xT∼N(0,I)x_T \sim \mathcal{N}(0, I).

Canonical Papers

Consistency Models

Song et al.2023ICML
Read paper →

Improved Techniques for Training Consistency Models

Song & Dhariwal2023arXiv
Read paper →

Connections

Prerequisites

∂Diffusion∇pScore Matching

Next Moves

Explore this concept from different angles — like a mathematician would.