GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models

GLIDE investigates diffusion models for text-conditional image synthesis, building on the observation that diffusion models generate high-quality images especially when paired with a guidance technique that trades off diversity for fidelity. The authors compare two guidance strategies, CLIP guidance and classifier-free guidance, evaluating which produces more photorealistic images and better caption alignment. They train a 3.5 billion parameter text-conditional diffusion model and additionally show the model can be fine-tuned to perform image inpainting for powerful text-driven image editing.

Human evaluators prefer classifier-free guidance over CLIP guidance for both photorealism and caption similarity, and often find its samples photorealistic. Notably, samples from the classifier-free GLIDE model are favored over those from DALL-E even when DALL-E uses expensive CLIP reranking. The authors train a smaller model on a filtered dataset and release its code and weights, demonstrating text-guided diffusion for both image generation and editing.