Bokeh Diffusion: Defocus Blur Control in Text-to-Image Diffusion Models
Bokeh Diffusion introduces a framework that allows precise control over defocus blur in text-to-image models by explicitly conditioning them on a physical blur parameter. It overcomes limitations of prompt engineering by using a hybrid training pipeline combining real images with synthetic augmentations. The core innovation is a grounded self-attention mechanism trained to ensure that changing the blur level preserves the original scene content, leading to flexible, lens-like bokeh control.