When diffusion models topped GANs on ImageNet

What it does

This repository holds the code and pre-trained weights for OpenAI’s 2021 paper Diffusion Models Beat GANs on Image Synthesis. It extends the earlier improved-diffusion codebase with classifier conditioning and architecture changes, shipping a full zoo of checkpoints for ImageNet at resolutions from 64×64 to 512×512, plus LSUN bedroom, cat, and horse models.

The interesting bit

Rather than baking class knowledge solely into the diffusion model, the system keeps a separate classifier on hand to nudge the denoising trajectory at inference time. You can tune the guidance strength, disable it entirely, or chain an upsampler to turn small samples into 256×256 or 512×512 images with sharper FID scores.

Key highlights

• Pre-trained ImageNet diffusion and classifier pairs at 64×64, 128×128, 256×256, and 512×512
• Classifier guidance via classifier_sample.py, with a configurable --classifier_scale to trade off fidelity and diversity
• Super-resolution upsamplers (64→256 and 128→512) that refine low-res outputs
• Sampling supports both fast DDIM (25 steps) and standard schedules up to 1000 timesteps
• Distributed classifier training scripts using MPI, with published hyperparameters for reproduction

Caveats

• Training documentation for diffusion models lives in the parent repository; this README only covers classifier training and sampling flags
• The documentation is essentially a quickstart manual of shell flags and download links, with little discussion of model architecture or theory

Verdict A useful historical artifact and reproduction target for generative-model researchers. If you need a modern, maintained diffusion library, this is not it.