Controllable Image Generation

The field studies how to make image generators follow specific instructions about what an image contains and how its elements look and are arranged. Common approaches condition diffusion or flow models on layouts, boxes, sketches, masks, poses, depth maps, and reference images, or steer their attention, latent representations, noise, and activations; newer systems also plan and check images through iterative reasoning. Researchers evaluate control using compositional and spatial benchmarks as well as domain-specific tasks such as medical imaging, fashion, and remote sensing. A central challenge is meeting precise local constraints while preserving realism, identity, prompt meaning, and diversity, especially when several controls interact.