icon Generative Refocusing
Flexible Defocus Control from a Single Image

Chun-Wei Tuan Mu1 Cheng-De Fan1 Jia-Bin Huang2 Yu-Lun Liu1
1National Yang Ming Chiao Tung University 2University of Maryland, College Park
arXiv Video HF Demo Code

Given a single input image, our method enables flexible refocusing and aperture adjustment.
The heatmap overlay visualizes the synthesized defocus map.

Historical Images

Input
Ours
◀▶
Input
Ours
Input
Ours
◀▶
Input
Ours

In-the-wild Images

Input
Ours
◀▶
Input
Ours
Input
Ours
◀▶
Input
Ours

Interactive Refocusing

Input Image
Input Image
Generative refocusing demo
Click on a subject to refocus
Small Large

Comparison with Gemini 3 (Nano Banana Pro)

Input
Input
Nano Banana Pro
Nano Banana Pro
Ours
Ours

Method

Our method decomposes single-image refocusing into two stages: (a) Defocus Deblurring and (b) Shallow Depth-of-Field Synthesis. Given a blurry input image Iin, DeblurNet restores a high-quality all-in-focus image Iaif. In the second stage, BokehNet takes the all-in-focus image Iaif, the defocus map Ddef, and optionally a specific aperture shape A as inputs to synthesize the refocused output Iout.
Method Pipeline

Visual Comparison

Case:
Method:
Method
Ours
◀▶
Method
Ours
Case:
Method:
Method
Ours
◀▶
Method
Ours
Case:
Method:
Method
Ours
◀▶
Method
Ours

BibTeX

@article{Genfocus2025,
          title     = {Generative Refocusing: Flexible Defocus Control from a Single Image},
          author    = {Tuan Mu, Chun-Wei and Fan, Cheng-De and Huang, Jia-Bin and Liu, Yu-Lun},
          journal   = {arXiv preprint arXiv:2512.16923},
          year      = {2025}
        }