[KongchangAI]
Model

FlashReFocus

A single-image refocusing method that decouples all-in-focus image restoration from blur rendering. It uses a one-step diffusion model to restore the all-in-focus image and a lightweight defocus-biased attention renderer to render blur in real time.

Timeline (last 90 days)

Oct 19

FlashReFocus decouples all-in-focus image restoration from blur rendering, so the diffusion model runs only once, and subsequent interactive editing is completed in real time by a lightweight network.

Unverified50%
Oct 19

On the EBB! and 3CReal benchmarks, FlashReFocus achieves the best perceptual fidelity in the zero-shot bokeh rendering task.

Unverified50%
Oct 19

Compared with GenRefocus, FlashReFocus refocuses 148.3 times faster.

Unverified50%
Oct 19

Each subsequent edit in FlashReFocus takes only about half a second, a speedup of up to 462 times compared with GenRefocus.

Unverified50%
Oct 19

The current FlashReFocus method still depends on the accuracy of depth estimation, and its performance in complex scenes such as transparent objects and fine hair remains to be further verified.

Unverified50%
Oct 19

By editing the defocus map, FlashReFocus can achieve advanced controls such as tilt-shift effects or dual focal planes.

Unverified50%

All Facts (6)

Source Articles