Diffusion-Based Image Generation
Overview
Diffusion models learn to turn noise into images through a sequence of denoising steps, often guided by text, labels, sketches, layouts, or reference images. Work spans pixel-space and latent-space models, U-Net and Transformer architectures, and related flow-matching methods, with efforts to improve training objectives, representations, and controllability. Researchers evaluate quality and prompt or spatial alignment on benchmarks such as ImageNet and COCO, and apply the models to editing, medical imaging, remote sensing, and synthetic data generation. A central challenge is balancing image quality, diversity, precise control, and reliability against the computation required for training and sampling; synthetic images also need validation for downstream usefulness and safety.
Latest papers
328 papers46 papers in the last four weeks, up 171% on the four weeks before. 0.3% of all new papers.