OpenAI Consistency Models: Fast One-Step Generative AI

OpenAI has introduced Consistency Models, a new family of generative generative models designed to accelerate generation speed. By directly mapping noise to data, these models enable high-quality, one-step generation, addressing the slow iterative sampling process required by traditional diffusion models.

Technical Approach: Direct Noise-to-Data Mapping

Consistency Models map noise to data directly, allowing for fast one-step generation by design. Unlike diffusion models, which require multiple iterations to refine a sample, Consistency Models can produce a high-quality output in a single step. However, they remain flexible; they models support multistep sampling, allowing users to trade increased compute for higher sample quality.

Training Methods: Distillation and Standalone Training

Consistency Models can be deployed using two distinct training pathways:

  1. Distillation: They can be trained by distilling pre-trained diffusion models, which leverages existing diffusion-based knowledge to accelerate generation.
  2. Standalone Training: They can be trained as independent generative models from scratch, without requiring a pre-existing diffusion model.

Capabilities and Zero-Shot Data Editing

Beyond simple generation, Consistency Models support zero-shot data editing tasks without requiring explicit training for these specific functions. These capabilities include:

  • Image Inpainting: Filling in missing parts of an image.
  • Colorization: Adding color to grayscale images.
  • **Super-resolution: Increasing the resolution of an image.

Performance Benchmarks

Consistency Models outperform existing distillation techniques for diffusion models in one- and few-step sampling. When trained as standalone models, they outperform existing one-step, non-adversarial generative models on standard benchmarks.

Key performance metrics include the following state-of-the-art FID scores for one-step generation:

  • CIFAR-10: 3.55 FID
  • ImageNet 64x64: 6.20 FID

Experiments were conducted across standard benchmarks including CIFAR-10, ImageNet 64x64, and LSUN 256x256.

Sources