Attention Control in Diffusion Models

The field studies how self-, cross-, and joint-attention route information between text, image or video regions, references, and denoising steps. Common approaches modify attention maps, logits, or values at inference; use masks and spatial routing to bind tokens to regions; share reference keys and values; or adapt attention and guidance across denoising steps, sometimes with training-time supervision. These controls target tasks such as multi-object composition, localized editing, style transfer, subject preservation, and video consistency, with evaluations using benchmarks such as T2I-CompBench, PIE-Bench, CountBench, and VBench. A central tension is that attention interventions can improve alignment or local control but may have entangled, distributed effects, making preservation and quality gains inconsistent across models and settings.