RAW images captured by different camera sensors exhibit substantial domain shifts due to varying spectral responses, noise characteristics, and tone behaviors, complicating their direct use in downstream computer vision tasks. Prior methods address this problem by training domain-specific RAW-to-RAW translators for each source-target pair, but such approaches do not scale to real-world scenarios involving multiple types of commercial cameras. In this work, we introduce MERIT, the first unified framework for multi-domain RAW image translation, which leverages a single model to perform translations across arbitrary camera domains. To address domain-specific noise discrepancies, we propose a sensor-aware noise modeling loss that explicitly aligns the signal-dependent noise statistics of the generated images with those of the target domain. We further enhance the generator with a conditional multi-scale large kernel attention module for improved context and sensor-aware feature modeling. To facilitate standardized evaluation, we introduce MDRAW, the first dataset tailored for multi-domain RAW image translation, comprising both paired and unpaired RAW captures from five diverse camera sensors across a wide range of scenes. Extensive experiments demonstrate that MERIT outperforms prior models in both quality (5.56 dB improvement) and scalability (80% reduction in training iterations).
Figures & tables
Figure 2: MERIT Framework Overview. (a). Overall architecture of MERIT. (b). The style encoder E extracts domain-specific embeddings from target RAW exemplars using a transformer-based architecture. (c). The generator G synthesizes target-domain RAW images. (d). The discriminator D operates on image patches and predicts the realism of each patch, using majority voting to make final decisions. (e). The proposed MS-KLA module captures multi-scale features and modulates them via style-aware channel attention.
Figure 3: Visualization of the sensor-aware noise modeling process. (a). Input RAW image. (b). Gradient magnitude map computed via the Sobel operator. (c). Red-shaded regions indicate low-texture patches selected for sensor noise profile estimation.
Camera
Sensor
No. of unpaired images
Samsung Galaxy S23 Ultra
ISOCELL HP2
81
Huawei P30
Sony IMX650 Exmor RS
119
iPhone 13 Pro Max
Sony IMX703
98
Nikon Z5
Sony IMX128
114
Canon EOS Rebel T6
DIGIC 4+
107
Table 1: Summarization of our collected dataset (MDRAW).
Figure 4: Qualitative results on the RAW-to-RAW mapping dataset. Each group of images (from left to right) shows: the input source RAW image, predictions from CUT [ 26 ] , Xie et al . [ 33 ] , CycleGAN [ 37 ] , UVCGAN [ 30 ] , Rawformer [ 27 ] , our method MERIT, and the target ground truth. Odd-numbered rows display the translated RAW outputs, while even-numbered rows visualize the corresponding absolute error maps with respect to the ground truth.
Training Paradigm
Model
Samsung-S9 → iPhone-X
iPhone-X → Samsung-S9
PSNR ↑
SSIM ↑
MAE ↓
PSNR ↑
SSIM ↑
MAE ↓
Non‐learning
Global calibration (3 × 3) [ 24 ]
24.52
0.71
0.049
17.03
0.51
0.160
Global calibration (poly) [ 24 ]
24.88
0.72
0.048
16.88
0.50
0.160
FDA [ 35 ]
20.95
0.48
0.060
19.18
0.47
0.090
Semi‐supervised
Afifi et al. [ 1 ]
29.65
0.89
0.027
28.58
0.90
0.033
Unsupervised
CycleGAN [ 37 ]
24.55
0.76
0.046
25.21
0.76
0.042
Table 2: Quantitative results on the RAW‐to‐RAW mapping dataset. Best results are bolded and second-best are underlined .
Table 3: Cross-domain RAW2RAW translation results on MDRAW. Each cell reports results for translation from a source domain (column) to a target domain (row). Within each cell, three lines correspond to different methods: UVCGAN [ 30 ] , Xie et al. [ 33 ] , and MERIT (Ours) . Each line contains four metrics in the order of MAE ( ↓ )/ PSNR ( ↑ )/ SSIM ( ↑ )/ KL Divergence ( ↓ ).
Setting
Source→Target
MAE ↓
PSNR ↑
SSIM ↑
Baseline
iPhone-X → Samsung-S9
0.0256
30.99
0.8357
Samsung-S9 → iPhone-X
0.0186
33.75
0.8841
Avg all
0.0221
32.37
0.8599
+ SANM
iPhone-X → Samsung-S9
0.0250
31.24
0.8426
Table 4: Ablation study on RAW-to-RAW mapping dataset.
Figure 5: Sample images from our MDRAW benchmark. (a). RAW images captured under various lighting and scenes from five different camera sensors. (b). Example of aligned multi-domain RAW captures of the same scene, used for evaluation in cross-domain RAW2RAW translation.
Figure 6: Visualization of more captured images in MDRAW.
Figure 7: More qualitative results on the RAW-to-RAW mapping dataset.
No. of Domains
Method
Parameters (M)
Training Iteration (K) / Training Time (Minutes)
Avg MAE ↓
Avg PSNR ↑
Avg SSIM ↑
Avg KL Div ↓
3
UVCGAN [ 30 ]
186
228 / 1955
0.039
28.03
0.72
2.72
Xie et al . [ 33 ]
33.6
266 / 1338
0.038
27.82
0.73
2.33
MERIT (Ours)
58.7
120 / 1400
0.031
29.60
0.76
1.76
4
UVCGAN [ 30 ]
372
476 / 3956
0.038
27.85
0.72
2.66
Xie et al . [ 33 ]
67.2
507 / 2055
0.037
28.31
0.72
2.35
MERIT (Ours)
58.7
165 / 1925
0.033
29.08
0.74
2.07
Table 5: Comparison of different models under increasing numbers of domains. MERIT demonstrates both competitive performance and efficiency.
Table 6: 4 × 4 Cross-domain RAW2RAW translation results on MDRAW. Each cell reports results for translation from a source domain (column) to a target domain (row). Each line contains four metrics in the order of MAE ( ↓ )/ PSNR ( ↑ )/ SSIM ( ↑ )/ KL Divergence ( ↓ ).
Table 7: 3 × 3 Cross-domain RAW2RAW translation results on MDRAW. Each cell reports results for translation from a source domain (column) to a target domain (row). Each line contains four metrics in the order of MAE ( ↓ )/ PSNR ( ↑ )/ SSIM ( ↑ )/ KL Divergence ( ↓ ).
Loss Term
Setting
MAE ↓
PSNR ↑
SSIM ↑
Lnoise
λ=0.25
0.0203
32.81
0.08668
λ=0.75
0.0199
33.27
0.8707
λ=1
0.0185
33.60
0.8806
λ=1.5
0.0195
33.24
0.8769
λ=5
0.0212
32.36
0.8576
Lcycle−SSIM
λ=0.02
0.0222
32.21
0.8560
Table 8: Ablation study on loss weight sensitivity. We evaluate the impact of loss weighting coefficients λ for the proposed noise modeling loss Lnoise and the cycle consistency SSIM loss Lcycle-SSIM . Each cell reports average performance on the RAW-to-RAW mapping dataset. Best results for each loss term are highlighted in bold.
School of Electrical and Information Engineering, Tianjin University · State key laboratory of Smart Power Distribution Equipment and System, Tianjin University · Individual