Refocusing a single image after capture is inherently interactive: users click different focal points and adjust blur strength, expecting each edit to be quick. Recent generative methods, however, rerun large diffusion models for every edit, which is slow. The FlashReFocus authors argue that the task splits into two sub-problems: recovering an all-in-focus image is ill-posed and needs a generative prior, while rendering blur is largely determined by depth and the focal plane.

FlashReFocus therefore runs diffusion once per photograph to deblur and recover detail, then uses a lightweight renderer for subsequent edits. The renderer is conditioned on an explicit defocus map and uses defocus-bias attention, so pixels at similar defocus share information while pixels at other depths do not interfere. This keeps in-focus content unchanged and allows effects like tilt-shift or two focal planes by editing the map.

The authors also collected 3CReal, a benchmark of 89 image pairs from three cameras, to test generalization to unseen cameras. On EBB! and 3CReal, FlashReFocus achieves the best perceptual fidelity for zero-shot bokeh rendering. Compared with GenRefocus, it refocuses 148.3x faster and takes about half a second per further edit, a 462x speedup.