3CReal: real bokeh pairs from three camera and lens systems

3CReal is a benchmark for bokeh rendering and refocusing on unseen optics. Each of its 89 pairs shows one scene twice from a fixed tripod: a deep-depth-of-field photograph (the input) and a shallow-depth-of-field photograph of the same view (the target), with the point the target is focused on. Existing real bokeh benchmarks are each captured with a single camera and lens, so a method can fit the look of one lens without this being measured; 3CReal spans three camera and lens systems that none of the compared methods saw in training.

Collected by Ching-Heng Cheng, Chia-Ming Lee, Ming-Ching Chang, Xin Li, Yu-Lun Liu and Chih-Chung Hsu (National Cheng Kung University, State University of New York at Albany, National Yang Ming Chiao Tung University). It was introduced with FlashReFocus (paper, code, demo).

Composition

Camera and lens	Pairs	Deep DoF (input)	Shallow DoF (target)	Size
Canon EOS 750D, EF-S 18–55 mm f/3.5–5.6 IS STM	31	f/22–f/36	f/3.5–f/5.6	2966×1974
FUJIFILM X-E5, Viltrox AF 35 mm F1.7 Air	31	f/16	f/1.7–f/5.6	1900×1256 (2 portrait)
FUJIFILM X100VI, Fujinon 23 mm F2 (built-in)	27	f/16	f/2.0	1916×1272 (1 portrait), 2032×1135 (8 pairs, 16:9)
Total	89

Apertures come from the EXIF of the raw captures; the eight 16:9 X100VI pairs have no EXIF, so their exposure fields are empty. Scenes are on a university campus, with automatic exposure for each shot.

Collection and alignment

Each pair was taken from a tripod, changing only the aperture. Residual shifts from tripod vibration, scene motion or manual operation were corrected by ECC registration with a translation model (the input is warped onto the target), both images were cropped to their common area, and every aligned pair was inspected by hand. ECC removes global geometric shifts but not exposure differences, vignetting, independently moving objects (leaves, people, cars) or aperture-dependent occlusion, so full-reference scores should be read together with visual comparisons.

Files

original/<key>.jpg     deep depth of field: the input (keys 0-88)
bokeh/<key>.jpg        shallow depth of field: the target
focus_point.json       focal point of each target, in pixels (x right, y down, origin top-left)
metadata.csv           per pair: camera, lens, width, height, focal point, focal length, f-number, exposure time
                       and ISO of both shots, ECC shift (pixels) and ECC score
data/*.parquet         the same pairs with embedded images, for datasets.load_dataset and the viewer

The images are the files used in the paper (not re-encoded), except for the masked license plates in pairs 48 and 60 (see Privacy).

Usage

from datasets import load_dataset

ds = load_dataset("ming0531/3CReal", split="validation")
pair = ds[0]
pair["input"], pair["target"]              # PIL images: deep and shallow depth of field
pair["focal_x"], pair["focal_y"]           # where the target is focused
pair["camera"], pair["deep_f_number"], pair["shallow_f_number"]

or download the folders (huggingface-cli download ming0531/3CReal --repo-type dataset --local-dir 3CReal) and read original/, bokeh/ and focus_point.json directly; this is the layout the FlashReFocus evaluation scripts expect.

Evaluation protocol (FlashReFocus paper)

Every method renders the input at the annotated focal point; its blur strength (or blur parameter) is searched per image for the highest SSIM to the target. Reported metrics are PSNR, SSIM, MS-SSIM, LPIPS (AlexNet and VGG), DISTS, CLIP-I, CLIP-IQA, MUSIQ and MANIQA; see eval/bokeh_strength_search.py and eval/bokeh_table_metrics.py in the code repository.

Privacy

The photographs show public campus scenes. The two readable license plates (pairs 48 and 60) are blurred, in the input and the target alike; the paper's scores were computed before this masking, which changes FlashReFocus' mean PSNR on 3CReal by less than 0.05 dB. If you appear in an image, or find content that should not be shared, please open a discussion on this page and we will remove or mask it.

License

Creative Commons Attribution-NonCommercial 4.0: free to use, share and adapt for non-commercial purposes, with attribution (please cite the paper below).

Citation

@misc{cheng2026flashrefocus,
  title  = {FlashReFocus: Interactive Image Refocusing in Seconds},
  author = {Cheng, Ching-Heng and Lee, Chia-Ming and Chang, Ming-Ching and
            Li, Xin and Liu, Yu-Lun and Hsu, Chih-Chung},
  year   = {2026},
  url    = {https://ming053l.github.io/FlashReFocus/}
}