Bridging RGB and RAW: Single-step Deterministic Flow with Homogeneous Representation Alignment

Diedong Feng, Peiyi Zeng, Zhen Liu, Zhongyang Li, Bing Zeng, Shuaicheng Liu
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:30531-30551, 2026.

Abstract

Reconstructing high-fidelity RAW sensor data from processed RGB images is a fundamental yet ill-posed problem, plagued by irreversible information loss and complex non-linear ISP transformations. While generative models offer high-quality reconstruction, they suffer from prohibitive computational costs. Conversely, dominant regression-based methods are fast but susceptible to incoherent observational deviations, often yielding over-smoothed predictions that drift from the authentic signal manifold. To reconcile this trade-off, we propose SHADE, a Single-step Homogeneous Aligned DEterministic flow framework. We validate that, unlike point-to-point regression, the single-step deterministic flow captures global transport trends and enables intrinsic robustness against input perturbations. Furthermore, we introduce Homogeneous Representation Alignment to maximize fidelity. By leveraging a homogeneously initialized student-teacher DINO pair, this mechanism enforces alignment within a shared feature space, significantly amplifying the representational capacity. Extensive experiments demonstrate that SHADE achieves state-of-the-art performance on multiple benchmarks, establishing a new paradigm for accurate and efficient sensor data reconstruction.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-feng26w, title = {Bridging {RGB} and {RAW}: Single-step Deterministic Flow with Homogeneous Representation Alignment}, author = {Feng, Diedong and Zeng, Peiyi and Liu, Zhen and Li, Zhongyang and Zeng, Bing and Liu, Shuaicheng}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {30531--30551}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/feng26w/feng26w.pdf}, url = {https://proceedings.mlr.press/v306/feng26w.html}, abstract = {Reconstructing high-fidelity RAW sensor data from processed RGB images is a fundamental yet ill-posed problem, plagued by irreversible information loss and complex non-linear ISP transformations. While generative models offer high-quality reconstruction, they suffer from prohibitive computational costs. Conversely, dominant regression-based methods are fast but susceptible to incoherent observational deviations, often yielding over-smoothed predictions that drift from the authentic signal manifold. To reconcile this trade-off, we propose SHADE, a Single-step Homogeneous Aligned DEterministic flow framework. We validate that, unlike point-to-point regression, the single-step deterministic flow captures global transport trends and enables intrinsic robustness against input perturbations. Furthermore, we introduce Homogeneous Representation Alignment to maximize fidelity. By leveraging a homogeneously initialized student-teacher DINO pair, this mechanism enforces alignment within a shared feature space, significantly amplifying the representational capacity. Extensive experiments demonstrate that SHADE achieves state-of-the-art performance on multiple benchmarks, establishing a new paradigm for accurate and efficient sensor data reconstruction.} }
Endnote
%0 Conference Paper %T Bridging RGB and RAW: Single-step Deterministic Flow with Homogeneous Representation Alignment %A Diedong Feng %A Peiyi Zeng %A Zhen Liu %A Zhongyang Li %A Bing Zeng %A Shuaicheng Liu %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-feng26w %I PMLR %P 30531--30551 %U https://proceedings.mlr.press/v306/feng26w.html %V 306 %X Reconstructing high-fidelity RAW sensor data from processed RGB images is a fundamental yet ill-posed problem, plagued by irreversible information loss and complex non-linear ISP transformations. While generative models offer high-quality reconstruction, they suffer from prohibitive computational costs. Conversely, dominant regression-based methods are fast but susceptible to incoherent observational deviations, often yielding over-smoothed predictions that drift from the authentic signal manifold. To reconcile this trade-off, we propose SHADE, a Single-step Homogeneous Aligned DEterministic flow framework. We validate that, unlike point-to-point regression, the single-step deterministic flow captures global transport trends and enables intrinsic robustness against input perturbations. Furthermore, we introduce Homogeneous Representation Alignment to maximize fidelity. By leveraging a homogeneously initialized student-teacher DINO pair, this mechanism enforces alignment within a shared feature space, significantly amplifying the representational capacity. Extensive experiments demonstrate that SHADE achieves state-of-the-art performance on multiple benchmarks, establishing a new paradigm for accurate and efficient sensor data reconstruction.
APA
Feng, D., Zeng, P., Liu, Z., Li, Z., Zeng, B. & Liu, S.. (2026). Bridging RGB and RAW: Single-step Deterministic Flow with Homogeneous Representation Alignment. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:30531-30551 Available from https://proceedings.mlr.press/v306/feng26w.html.

Related Material