ECCV 2026

Camera × Display × Observer

Color Pass-Through via Camera-Display Coupling

Ruikang Li¹ Molin Li² Jiarui Wu¹ Zhe Wei³ Pengpeng Liu³ Tianfan Xue^1,†

¹CUHK MMLab ²Zhejiang University ³Central Media Technology Institute, Huawei

^† Corresponding author

Paper Code (Coming Soon)

Make a smartphone transparent, like Apple Vision Pro.

A real scene, smartphone, and observers followed by comparisons of default camera, learned white balance, ColorChecker calibration, and our method.

(a)

Commercial smartphone

A strong black-box ISP with production-grade white balance and color transforms.

(b)

Multi-illuminant Auto-WB

A state-of-the-art multi-illuminant white-balance method applied on top of (a) to further remove spatially varying illumination casts.

(c)

ColorChecker calibration

A strong conventional baseline using ground-truth color correspondences to map (a) into a predefined CIE reference color space.

(d)

Ours

NO illumination priors. NO pre-calibrated color space. For each specific device, we learn the complete camera-display path end to end.

+2.0 points on a 5-point user study

>2× improvement on quantitative metrics

1-step calibration for each observer

Quantitative evaluation

Large gains on both commercial phones Xiaomi 17 Pro Max / Huawei Pura 70 Pro

Evaluation uses 24 ColorChecker patches under ten color temperatures and five randomly sampled RGB-LED illuminants. Lower is better for ΔE and STRESS.

Method	PSNR ↑		ΔE mean ↓		STRESS ↓
Method	Huawei	Xiaomi	Huawei	Xiaomi	Huawei	Xiaomi
Default smartphone	13.78	14.61	14.62	13.49	26.23	25.07
ColorChecker calibration	15.02	16.36	18.49	15.32	38.85	36.42
Multi-illuminant Auto-WB	12.84	13.92	17.84	17.08	31.12	30.05
Ours	28.65	29.10	5.18	4.79	17.48	16.12

The core idea

One coupled path, from scene to display.

Conventional pipelines calibrate the camera and display independently, then connect them through a low-dimensional intermediate color space. We instead optimize the complete camera-to-display path for a specific observer.

Comparison between separated camera and display calibration and our end-to-end learned projector, alongside radiance and color pass-through objectives.

Exact radiance pass-through is generally impossible: a three-channel camera cannot preserve every spectral dimension of a real scene. Color pass-through asks a more practical question—can the displayed result look the same to the observer?

Method

Learn the projector. Correct what the camera cannot see.

Two complementary learned components transform a captured image into a display image that better matches the original scene for the target observer.

Camera-Display Projection

A lightweight pixel-wise neural projector learns the nonlinear end-to-end mapping of a paired camera and display from re-captured RGB data.

Camera-Null Color Correction

A learned predictor models the dominant metameric-black component, while a compact coefficient adapts the correction to each observer.

Ablation showing camera-display projection and camera-null correction for camera and non-camera observers.

Visual results

Closer color across unseen illumination.

Select an illumination setting to compare the default smartphone pipeline, per-scene ColorChecker calibration, and our method.

Direct in-scene comparison under indoor illumination.

Indoor illumination Unseen indoor illumination with spatially varying color casts.

Separated-view comparison

Displayed result versus the real scene.

Four scene comparisons showing default smartphone camera, learned white balance, ours, and the real-scene ground truth.

Citation

If this work helps your research, please cite us.

@inproceedings{li2026colorpassthrough,
  title     = {Color Pass-Through via Camera-Display Coupling},
  author    = {Li, Ruikang and Li, Molin and Wu, Jiarui and
               Wei, Zhe and Liu, Pengpeng and Xue, Tianfan},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}