|
|
| CARVIEW |
Select Language
Generative Refocusing
Demo
Historical Images
In-the-wild Images
Interactive Refocusing
Input Image
Click on a subject to refocus
Comparison with Gemini 3 (Nano Banana Pro)
Input
Nano Banana Pro
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, we optionally apply a pre-deblurring to obtain a conservative estimate Isyn,
then feed both Iin and Isyn into DeblurNet to recover 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 as inputs to synthesize the refocused output.
Visual Comparison
Case:
Method:
Case:
Method:
Case:
Method:
Visual Comparison for BokehNet Training Ablation
Training solely on synthetic data tends to result in an incorrect depth scale, leading to excessive bokeh
strength and noticeable artifacts along object boundaries.
Case:
Failure Cases
For severely blurred inputs, the model may hallucinate incorrect details (e.g., reconstructing "Jewelry shop"
as "Temnlry shop").
Related Links
There has been great works on refocusing or bokeh that were done concurrently by other researchers, feel free to check them out!