Text-Guided Image Editing

Text-guided image editing studies how to turn language instructions into targeted visual changes while keeping the source image’s identity, layout, and unrelated details intact. Common approaches adapt diffusion and flow models through inversion, attention or velocity guidance, spatial masks, and region localization; newer autoregressive editors and agents add explicit planning, tool use, and iterative feedback. Work also explores specialized settings such as scene text, portraits, products, documents, and multi-turn edits, with benchmarks increasingly testing localization, instruction reasoning, and human-perceived quality. The central tension remains balancing strong edits that satisfy complex intent against faithful preservation, especially for small targets, implicit constraints, and long editing sequences.