Consistency Models

Diffusion models are the workhorses behind most of the image, video, and audio generators people use today, but they have a dirty secret: they're slow. To turn noise into a coherent picture, a diffusion model has to run through dozens or even hundreds of denoising steps, each one a full pass through a neural network. That's fine for a research demo. It's less fine when you want real-time generation or you're paying for compute at scale.

OpenAI's answer is something it calls consistency models, and the core idea is almost stubbornly simple. Instead of learning to nudge noisy data toward slightly-less-noisy data over many small steps, a consistency model learns to map any point on that noisy trajectory directly back to the clean result in one jump. Train it right, and you get an approach that can generate a usable sample in a single pass through the network. No iterative chain, no waiting.

The trick isn't free, obviously. Consistency models can also be run with a handful of extra steps if you want to trade a bit of speed for a bit more quality, so it's less an all-or-nothing swap and more a dial. What makes it notable is that these models can either be distilled from an existing pretrained diffusion model or trained from scratch as standalone generators, which gives researchers flexibility depending on whether they already have a diffusion model sitting around or are starting fresh.

This matters because the iterative bottleneck has been one of the main practical complaints about diffusion-based generation since it took over from GANs. Shaving that down to one or a few steps, without gutting output quality, is the kind of unglamorous infrastructure win that quietly changes what's feasible to ship. Faster sampling means cheaper inference, and cheaper inference means generative tools that were previously batch-only novelties start looking like something you could bolt onto a live product.