GLIDE

OpenAI's diffusion model for photorealistic image synthesis and local editing via text prompts.

GLIDE (Guided Language-to-Image Diffusion for Generation and Editing) scales a 3.5 billion parameter diffusion model to outperform GANs and DALL-E in human preference benchmarks. It utilizes classifier-free guidance to maintain high spatial fidelity while adhering strictly to complex text descriptions. Beyond base generation, the architecture supports zero-shot image inpainting, allowing users to modify existing photos (changing a dog's breed or adding a sunset) by masking regions and providing natural language instructions. The model demonstrates that diffusion techniques offer superior photorealism and fine-grained control for professional creative workflows.

https://github.com/openai/glide-text2im