Why Low-Light Cameras Go Color Blind:
Removing Color Bias in Raw Denoising

ICCP 2026 · Recommended to Special Issue, PAMI 2026
Mohammad Mohammadi1,2,3 Sina Honari3 Stavros Tsogkas3 Tristan Aumentado-Armstrong3 Michael S. Brown3,4 Iqbal Mohomed3 Konstantinos G. Derpanis1,2,3,4 Alex Levinshtein3,* Igor Gilitschenski1,2,*
1University of Toronto 2Vector Institute 3Samsung AI Center Toronto 4York University *Joint supervision
arXiv Code Dataset Poster Twitter
Code & dataset — coming soon

TL;DR

We discover a systematic color shift in the SIDD ground-truth images (the purple tint). We propose a solution and introduce SIDD-CC, the color-corrected version of the dataset.

We also find that, in low light, the black level a camera reports in its metadata may be inaccurate, and this black-level error violates the zero-mean assumption behind raw denoising. As a result, the denoiser's output suffers from a color shift. We model the black-level error, supervise its estimation with a two-pass network, and erase the bias, setting a new state of the art among blind raw denoising methods.

SIDD-CC: SIDD with color-corrected ground truths

SIDD builds its "clean" ground truth by averaging many noisy frames — but a subtle mistake in the order of operations bakes a systematic color shift (a purple tint) into that ground truth. We diagnose it and release SIDD-CC, a corrected version.

SIDD-CC corrected ground truth, chess scene
SIDD ground truth, chess scene
SIDD SIDD-CC
SIDD-CC corrected ground truth, toy trucks
SIDD ground truth, toy trucks
SIDD SIDD-CC

Drag the divider — the left side is SIDD (purple tint), the right side is SIDD-CC (corrected).

Why: the order of operations

In a raw frame, a dark pixel's noise may dip below the sensor's black level. SIDD subtracts the black level and clips every frame before averaging — clipping discards those sub-zero values, so the average is pulled upward. This mean shift shows up as a color cast after rendering by ISP. Averaging first, then clipping, preserves the true mean.

SIDD Noisy frames→ Subtract black level & clip→ Average⟹ mean shift · purple tint
SIDD-CC Noisy frames→ Average→ Subtract black level & clip⟹ true colors

Black-level error estimation

The same problem returns at inference: the black level a camera reports in its metadata may be inaccurate. This black-level error violates the zero-mean assumption underlying Raw noise modeling, and a denoiser trained under that assumption bakes a color shift into its output. We model the black-level error explicitly, but a new noise model alone is not enough to solve this problem. Black-level error is a global distortion of the image, so the model used for denoising must be able to extract this distortion — and commonly used U-Net architectures, with their local receptive field, cannot capture global information. We therefore insert a U-Net encoder prior to denoising: a first model to estimate the black-level error from a global view of the noisy input, and a second model to denoise the noisy image using the corrected black level. An auxiliary loss supervises the estimate end-to-end, so the correction is learned rather than assumed.

Method overview: training and inference pipelines with black-level-error estimation and correction
Method overview. A first network — the BLE Predictor — estimates the black-level error (ℰ′bl) from a global view of the raw input; that estimate corrects the black-level subtraction during preprocessing before a second network, the Denoiser, cleans the image. Training (top): our noise synthesis injects a known black-level error ℰbl, an L1 loss supervises the prediction against it, and a second L1 loss supervises the denoised output. Inference (bottom): the predictor estimates the error directly from the noisy input.

Quantitative results

Pick a dataset and noise ratio to compare PSNR and SSIM across every method. Among blind methods Ours leads, and it stays competitive with calibration-based methods that require extra per-camera measurements.

Qualitative results

Choose a dataset and example, then toggle any method to compare it against the ground truth. Drag the divider to reveal each side. Ours removes the color shift that lingers in the blind baselines.

University of Toronto Vector Institute Samsung AI Center Toronto York University