Consistency Models

Consistency models are a family of generative models that can generate samples in one or a few steps, rather than the tens or hundreds of iterative denoising steps typical of diffusion models. Introduced by OpenAI researchers in March 2023, they learn a function that maps any point along a diffusion trajectory — a probability flow ODE — back to the same starting point, making all points on a trajectory "consistent" with one another.^[1]

Background

Diffusion and score-based models generate data by slowly removing noise over many steps, which gives high sample quality but slow inference: a state-of-the-art image model may need 25–1,000 network evaluations per sample. This latency matters for interactive applications such as real-time image editing, video generation and on-device synthesis. Earlier speed-up approaches — fewer-step samplers, distillation into adversarial networks (as in progressive distillation and GAN-based distillation methods) — either degraded quality or required discriminator training.^[1]

Consistency training sidesteps both problems. Because the probability flow ODE of a diffusion process has a known analytic solution mapping every noisy point to the same clean endpoint, a model can be trained to produce identical outputs from any two points on the same trajectory — a self-supervised objective needing no adversarial component.^[1]

How It Works

• Probability flow ODE. Diffusion forward noising can be written as an ordinary differential equation whose trajectories all converge to the same clean data point; time is integrated backward during sampling.
• Self-consistency constraint. The network fθ is constrained so that f(x_t, t) = f(x_s, s) whenever x_s and x_t lie on the same ODE trajectory — enforced by parameterising outputs around a fixed anchor (e.g. the model's prediction at t = 0) so the equality holds by construction.
• Two training routes. Consistency distillation starts from a pre-trained diffusion model and matches its trajectories; consistency training (CT) estimates trajectories directly from data using pairs of adjacent noisy samples, requiring no pre-trained model at all.^[1]
• Few-step trade-off. Sampling can stop after any number of steps; more steps trade compute for sample quality — a property called the consistency scaling law.^[2]

Results and Impact

The original paper reported state-of-the-art one-step FID of 3.55 on CIFAR-10 and 6.20 on ImageNet 64×64, outperforming prior diffusion distillation techniques; trained standalone, consistency models also outperformed existing one-step non-adversarial generators.^[1] The 2024 follow-up, Improved Techniques for Training Consistency Models, sharpened the training recipe (teacher averaging, EMA improvements) and further improved one- and few-step quality.^[2] The idea influenced later fast-sampling work, including latent consistency models for text-to-image, and sits alongside flow matching and diffusion distillation as the main toolkit for making generative models fast enough for production.

Applications

One- and few-step generation is critical for real-time video generation (see Sora, AI video generation), interactive image editing, mobile and edge inference where latency and battery dominate, and high-throughput synthetic-data pipelines where cost per sample scales with the number of steps.

>Key Takeaways

• Consistency models generate samples in one or a few steps by enforcing agreement across a diffusion ODE trajectory.
• They can be trained by distilling a pre-trained diffusion model or from scratch via consistency training — no adversarial training needed.
• They set one-step FID records at introduction (3.55 CIFAR-10) and enabled the few-step efficiency used in modern image and video tools.
• The core value in production is latency and cost: fewer network evaluations per sample.

See Also

• Diffusion Model
• Flow Matching
• Knowledge Distillation
• AI Video Generation
• Sora
• Inference