pith. sign in

arxiv: 2512.05859 · v2 · submitted 2025-12-05 · 💻 cs.CV

Edit-aware RAW Reconstruction

Abhijith Punnappurath , Luxi Zhao , Ke Zhao , Hue Nguyen , Radek Grzeszczuk , Michael S. Brown This is my paper

Pith reviewed 2026-05-17 00:37 UTC · model grok-4.3

classification 💻 cs.CV
keywords RAW reconstructionedit-aware lossdifferentiable ISPsRGB to RAWphotofinishing pipelineimage editingcamera image processing
0
0 comments X

The pith

A plug-and-play loss using a differentiable ISP makes RAW reconstructions more robust to post-capture edits.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces an edit-aware loss function that can be added to any RAW reconstruction method from sRGB images. It works by rendering both the ground-truth and reconstructed RAW through a modular differentiable ISP whose parameters are randomly sampled to mimic real photofinishing variations, then measuring the difference in the resulting sRGB outputs. This matters because users mostly edit sRGB JPEGs rather than storing and editing RAW files, so the recovered RAW needs to support those edits without quality loss. The loss is designed as a general add-on that improves fidelity across rendering styles and editing operations.

Core claim

The authors establish that incorporating their edit-aware loss, computed between ground-truth and reconstructed RAWs after both pass through a modular differentiable ISP with randomly sampled parameters modeling practical camera pipelines, improves sRGB reconstruction quality by up to 1.5-2 dB PSNR across various editing conditions. When applied to metadata-assisted RAW reconstruction methods, the same loss further enables fine-tuning for target edits and yields additional gains.

What carries the argument

Modular differentiable image signal processor (ISP) that renders reconstructed and ground-truth RAWs through randomly sampled photofinishing parameters so the loss can be evaluated in sRGB space.

If this is right

  • Recovered RAWs become more robust to diverse rendering styles and editing operations performed on sRGB outputs.
  • Metadata-assisted RAW reconstruction methods gain the ability to fine-tune specifically for target edits with further quality gains.
  • Existing RAW reconstruction frameworks receive a general mechanism for enhancing edit fidelity and rendering flexibility without redesigning the core model.
  • Consumer workflows improve because photographic editing, the main reason for wanting RAW data, can be performed more accurately after reconstruction.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same loss could be tested on collections of real user edits to check whether the simulated parameter distributions match actual post-capture behavior.
  • Similar differentiable pipeline modeling might transfer to related tasks such as denoising or tone mapping where ISP variations also affect final image quality.
  • The approach opens the possibility of training reconstructors that are conditioned on specific editing intents rather than treating all edits equally.

Load-bearing premise

The modular differentiable ISP with randomly sampled parameters from carefully designed distributions sufficiently models the space of real-world photofinishing pipelines and post-capture edits that users actually apply.

What would settle it

Measuring whether the reported 1.5-2 dB PSNR gain in sRGB space disappears when the same reconstructed RAWs are passed through a set of real camera ISPs or user-applied edits that fall outside the training parameter distributions.

Figures

Figures reproduced from arXiv: 2512.05859 by Abhijith Punnappurath, Hue Nguyen, Ke Zhao, Luxi Zhao, Michael S. Brown, Radek Grzeszczuk.

Figure 1
Figure 1. Figure 1: (A) Existing RAW reconstruction methods typically em [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our edit-aware loss framework and differentiable ISP design. (A) Conventional RAW reconstruction models optimize [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results of adding our edit-aware loss to the RAW reconstruction method in CAM [ [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results of adding our edit-aware loss to the RAW recovery method in RAW Diffusion [ [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results of fine-tuning (FT) a UNet-based [ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
read the original abstract

Users frequently edit camera images post-capture to achieve their preferred photofinishing style. While editing in the RAW domain provides greater accuracy and flexibility, most edits are performed on the camera's display-referred output (e.g., 8-bit sRGB JPEG) since RAW images are rarely stored. Existing RAW reconstruction methods can recover RAW data from sRGB images, but these approaches are typically optimized for pixel-wise RAW reconstruction fidelity and tend to degrade under diverse rendering styles and editing operations. We introduce a plug-and-play, edit-aware loss function that can be integrated into any existing RAW reconstruction framework to make the recovered RAWs more robust to different rendering styles and edits. Our loss formulation incorporates a modular, differentiable image signal processor (ISP) that simulates realistic photofinishing pipelines with tunable parameters. During training, parameters for each ISP module are randomly sampled from carefully designed distributions that model practical variations in real camera processing. The loss is then computed in sRGB space between ground-truth and reconstructed RAWs rendered through this differentiable ISP. Incorporating our loss improves sRGB reconstruction quality by up to 1.5-2 dB PSNR across various editing conditions. Moreover, when applied to metadata-assisted RAW reconstruction methods, our approach enables fine-tuning for target edits, yielding further gains. Since photographic editing is the primary motivation for RAW reconstruction in consumer imaging, our simple yet effective loss function provides a general mechanism for enhancing edit fidelity and rendering flexibility across existing methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a plug-and-play edit-aware loss for RAW reconstruction from sRGB images. It employs a modular differentiable ISP with randomly sampled parameters from carefully designed distributions to simulate photofinishing pipelines and post-capture edits. The loss is computed in sRGB space between renderings of ground-truth and reconstructed RAW images through the same ISP. The authors claim this improves sRGB reconstruction quality by up to 1.5-2 dB PSNR across editing conditions and enables further gains via fine-tuning when applied to metadata-assisted RAW reconstruction methods.

Significance. If the central assumption holds, the work has practical significance for consumer imaging applications where RAW reconstruction is motivated by editing flexibility rather than pixel-wise fidelity. The plug-and-play design allows integration into existing frameworks without architectural changes, and the focus on edit robustness is a clear strength. Credit is given for the modular ISP formulation and the emphasis on simulating realistic parameter variations during training.

major comments (2)
  1. [Method (ISP module and parameter sampling)] The reported 1.5-2 dB PSNR gains and robustness claims rest on the modular differentiable ISP's randomly sampled parameter distributions faithfully covering real user edits and photofinishing pipelines. The manuscript should provide explicit validation (e.g., statistical comparison of sampled parameters against real edited image datasets or outputs from commercial editors) to confirm the distributions are representative rather than ad-hoc.
  2. [Experiments] Experiments section: the abstract states improvements 'across various editing conditions' and 'further gains' for metadata-assisted methods. The paper must detail the exact baselines, whether gains persist against stronger or more recent RAW reconstruction methods, and results on real (non-simulated) edits to ensure the improvements are not limited to the training distribution.
minor comments (2)
  1. [Abstract] Abstract: the phrase 'up to 1.5-2 dB' should be supplemented with average gains and standard deviations in the main results to allow precise assessment of the improvement magnitude.
  2. [Method] Ensure all ISP module parameters (tone mapping, color correction, sharpening, etc.) are explicitly listed with their sampling ranges and any correlations modeled.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review. We address each major comment below and describe the revisions we will make to improve the manuscript.

read point-by-point responses

  1. Referee: [Method (ISP module and parameter sampling)] The reported 1.5-2 dB PSNR gains and robustness claims rest on the modular differentiable ISP's randomly sampled parameter distributions faithfully covering real user edits and photofinishing pipelines. The manuscript should provide explicit validation (e.g., statistical comparison of sampled parameters against real edited image datasets or outputs from commercial editors) to confirm the distributions are representative rather than ad-hoc.

    Authors: We agree that additional validation would strengthen the presentation. The parameter distributions were constructed from ranges and variations documented in prior ISP and camera pipeline literature to reflect practical photofinishing operations. In the revised manuscript we will add an explicit validation subsection (or appendix) that statistically compares the sampled parameter distributions against parameter statistics extracted from real edited images in public datasets and from commercial editors such as Adobe Lightroom and Photoshop, including distribution plots and quantitative metrics. revision: yes

  2. Referee: [Experiments] Experiments section: the abstract states improvements 'across various editing conditions' and 'further gains' for metadata-assisted methods. The paper must detail the exact baselines, whether gains persist against stronger or more recent RAW reconstruction methods, and results on real (non-simulated) edits to ensure the improvements are not limited to the training distribution.

    Authors: We will revise the Experiments section to list all baselines with full citations and implementation details. We will expand the comparison set to include additional recent RAW reconstruction methods and report whether the observed gains hold. For real (non-simulated) edits, we will add quantitative results on a collection of real user-edited images drawn from available public sources, while noting that ground-truth RAW is obtained via the same controlled capture protocol used for the simulated data; this will demonstrate generalization beyond the training distribution. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper defines an edit-aware loss via a modular differentiable ISP whose module parameters are randomly sampled from designed distributions during training; the loss measures sRGB-space difference between ground-truth and reconstructed RAW images after identical rendering. The reported 1.5-2 dB PSNR gains are presented as empirical outcomes on sRGB reconstruction quality under various editing conditions, not as a mathematical identity or quantity forced by the same fitted parameters. No equation reduces the central result to its inputs by construction, no uniqueness theorem or self-citation is invoked as load-bearing justification, and the evaluation metric (PSNR) remains an external, standard benchmark independent of the training distributions. The derivation is therefore self-contained against external validation rather than tautological.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The method rests on the assumption that a modular differentiable ISP can be parameterized to cover realistic photofinishing variation; no new physical entities are introduced.

free parameters (1)
  • ISP module parameter distributions
    Distributions for tone curves, color matrices, sharpening, etc., are described as 'carefully designed' to model practical variations; these act as tunable free parameters that define the training distribution.
axioms (1)
  • domain assumption A modular differentiable ISP can faithfully simulate real camera photofinishing pipelines when its parameters are randomly sampled.
    Invoked in the loss formulation section of the abstract.

pith-pipeline@v0.9.0 · 5574 in / 1295 out tokens · 49490 ms · 2026-05-17T00:37:58.532067+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

  1. [1]

    Adobe Digital Negative (DNG).https : / / helpx . adobe . com / camera - raw / digital - negative . html, . Accessed: 2025-11-04. 6

    work page 2025

  2. [2]

    Accessed: 2025-11-04

    DNG Specification.https://helpx.adobe.com/ content / dam / help / en / photoshop / pdf / dng _ spec_1_6_0_0.pdf, . Accessed: 2025-11-04. 5

    work page 2025

  3. [3]

    CIE XYZ Net: Unprocessing images for low-level computer vision tasks.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(9):4688–4700, 2021

    Mahmoud Afifi, Abdelrahman Abdelhamed, Abdullah Abuolaim, Abhijith Punnappurath, and Michael S Brown. CIE XYZ Net: Unprocessing images for low-level computer vision tasks.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(9):4688–4700, 2021. 3, 8, 12, 13

    work page 2021

  4. [4]

    Time-aware auto white balance in mobile photography

    Mahmoud Afifi, Luxi Zhao, Abhijith Punnappurath, Mo- hamed A Abdelsalam, Ran Zhang, and Michael S Brown. Time-aware auto white balance in mobile photography. In ICCV, 2025. 6, 7, 8, 11, 12, 13

    work page 2025

  5. [5]

    ReRAW: RGB-to-RAW image recon- struction via stratified sampling for efficient object detection on the edge

    Radu Berdan, Beril Besbinar, Christoph Reinders, Junji Ot- suka, and Daisuke Iso. ReRAW: RGB-to-RAW image recon- struction via stratified sampling for efficient object detection on the edge. InCVPR, 2025. 2, 3

    work page 2025

  6. [6]

    Unprocessing im- ages for learned raw denoising

    Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing im- ages for learned raw denoising. InCVPR, 2019. 3

    work page 2019

  7. [7]

    Ayan Chakrabarti, Daniel Scharstein, and Todd E. Zickler. An empirical camera model for internet color vision. In BMVC, 2009. 3

    work page 2009

  8. [8]

    Modeling radiometric uncertainty for vision with tone- mapped color images.IEEE Transactions on Pattern Analy- sis and Machine Intelligence, 36(11):2185–2198, 2014

    Ayan Chakrabarti, Ying Xiong, Baochen Sun, Trevor Dar- rell, Daniel Scharstein, Todd Zickler, and Kate Saenko. Modeling radiometric uncertainty for vision with tone- mapped color images.IEEE Transactions on Pattern Analy- sis and Machine Intelligence, 36(11):2185–2198, 2014. 3

    work page 2014

  9. [9]

    Model-based image signal processors via learnable dictionaries

    Marcos V Conde, Steven McDonagh, Matteo Maggioni, Ales Leonardis, and Eduardo P ´erez-Pellitero. Model-based image signal processors via learnable dictionaries. InAAAI,

    work page

  10. [10]

    Reversed image signal processing and RAW reconstruction

    Marcos V Conde, Radu Timofte, Yibin Huang, Jingyang Peng, Chang Chen, Cheng Li, Eduardo P´erez-Pellitero, Fen- glong Song, Furui Bai, Shuai Liu, et al. Reversed image signal processing and RAW reconstruction. AIM 2022 chal- lenge report. InECCV AIM Workshop, 2022. 3

    work page 2022

  11. [11]

    NILUT: Conditional neural implicit 3D lookup tables for image enhancement

    Marcos V Conde, Javier Vazquez-Corral, Michael S Brown, and Radu Timofte. NILUT: Conditional neural implicit 3D lookup tables for image enhancement. InAAAI, 2024. 5, 6

    work page 2024

  12. [12]

    Debevec and Jitendra Malik

    Paul E. Debevec and Jitendra Malik. Recovering high dy- namic range radiance maps from photographs. InACM SIG- GRAPH, 2008. 3

    work page 2008

  13. [13]

    Mobile computational photography: A tour.Annual review of vision science, 7(1):571–604, 2021

    Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Mobile computational photography: A tour.Annual review of vision science, 7(1):571–604, 2021. 1, 2

    work page 2021

  14. [14]

    Grossberg and Shree K

    Michael D. Grossberg and Shree K. Nayar. Determining the camera response from images: What is knowable?IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(11):1455–1467, 2003. 3

    work page 2003

  15. [15]

    A software platform for manipulating the camera imaging pipeline

    Hakki Can Karaimer and Michael S Brown. A software platform for manipulating the camera imaging pipeline. In ECCV, 2016. 5, 6

    work page 2016

  16. [16]

    Seon Joo Kim, Hai Ting Lin, Zheng Lu, Sabine S ¨usstrunk, Stephen Lin, and Michael S. Brown. A new in-camera imaging model for color computer vision and its application. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 34(12):2289–2302, 2012. 2, 3

    work page 2012

  17. [17]

    ParamISP: Learned forward and inverse ISPs using camera parameters

    Woohyeok Kim, Geonu Kim, Junyong Lee, Seungyong Lee, Seung-Hwan Baek, and Sunghyun Cho. ParamISP: Learned forward and inverse ISPs using camera parameters. InCVPR,

    work page

  18. [18]

    Adam: A method for stochastic optimization.ICLR, 2014

    Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization.ICLR, 2014. 9

    work page 2014

  19. [19]

    Metadata- based RAW reconstruction via implicit neural functions

    Leyi Li, Huijie Qiao, Qi Ye, and Qinmin Yang. Metadata- based RAW reconstruction via implicit neural functions. In CVPR, 2023. 2 13

    work page 2023

  20. [20]

    Nonuniform lattice regression for modeling the cam- era imaging pipeline

    Hai Ting Lin, Zheng Lu, Seon Joo Kim, and Michael S Brown. Nonuniform lattice regression for modeling the cam- era imaging pipeline. InECCV, 2012. 5

    work page 2012

  21. [21]

    Tomoo Mitsunaga and Shree K. Nayar. Radiometric self cal- ibration. InCVPR, 1999. 3

    work page 1999

  22. [22]

    Modelling the scene dependent imaging in cameras with a deep neural network

    Seonghyeon Nam and Seon Joo Kim. Modelling the scene dependent imaging in cameras with a deep neural network. InICCV, 2017. 3, 12

    work page 2017

  23. [23]

    Learning sRGB-to-raw- RGB de-rendering with content-aware metadata

    Seonghyeon Nam, Abhijith Punnappurath, Marcus A Brubaker, and Michael S Brown. Learning sRGB-to-raw- RGB de-rendering with content-aware metadata. InCVPR,

    work page

  24. [24]

    2, 3, 4, 6, 7, 8, 9, 10, 11, 12

    work page

  25. [25]

    RAW image re- construction using a self-contained sRGB-JPEG image with only 64 KB overhead

    Rang MH Nguyen and Michael S Brown. RAW image re- construction using a self-contained sRGB-JPEG image with only 64 KB overhead. InCVPR, 2016. 2

    work page 2016

  26. [26]

    RAW image re- construction using a self-contained sRGB-JPEG image with small memory overhead.International Journal of Computer Vision, 126(6):637–650, 2018

    Rang MH Nguyen and Michael S Brown. RAW image re- construction using a self-contained sRGB-JPEG image with small memory overhead.International Journal of Computer Vision, 126(6):637–650, 2018. 2

    work page 2018

  27. [27]

    Adaptive histogram equalization and its variations.Com- puter vision, graphics, and image processing, 39(3):355– 368, 1987

    Stephen M Pizer, E Philip Amburn, John D Austin, Robert Cromartie, Ari Geselowitz, Trey Greer, Bart ter Haar Romeny, John B Zimmerman, and Karel Zuiderveld. Adaptive histogram equalization and its variations.Com- puter vision, graphics, and image processing, 39(3):355– 368, 1987. 9

    work page 1987

  28. [28]

    Spatially aware metadata for raw reconstruction

    Abhijith Punnappurath and Michael S Brown. Spatially aware metadata for raw reconstruction. InWACV, 2021. 2

    work page 2021

  29. [29]

    RAW-diffusion: RGB-guided dif- fusion models for high-fidelity RAW image generation

    Christoph Reinders, Radu Berdan, Beril Besbinar, Junji Ot- suka, and Daisuke Iso. RAW-diffusion: RGB-guided dif- fusion models for high-fidelity RAW image generation. In WACV, 2025. 1, 2, 3, 5, 6, 7, 9, 10, 12

    work page 2025

  30. [30]

    Ronneberger, P.Fischer, and T

    O. Ronneberger, P.Fischer, and T. Brox. U-Net: Convolu- tional networks for biomedical image segmentation. InMed- ical Image Computing and Computer-Assisted Intervention (MICCAI), 2015. 6, 7, 8, 9, 11, 12

    work page 2015

  31. [31]

    CRC Press, 2nd edition, 2013

    Gaurav Sharma and Raja Bala.Digital Color Imaging Hand- book. CRC Press, 2nd edition, 2013. 7

    work page 2013

  32. [32]

    Raw image reconstruc- tion with learned compact metadata

    Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Raw image reconstruc- tion with learned compact metadata. InCVPR, 2023. 2, 3

    work page 2023

  33. [33]

    Beyond learned metadata-based raw image reconstruction.International Journal of Computer Vision, 132(12):5514–5533, 2024

    Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap- Pui Chau, Alex C Kot, and Bihan Wen. Beyond learned metadata-based raw image reconstruction.International Journal of Computer Vision, 132(12):5514–5533, 2024. 2

    work page 2024

  34. [34]

    Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004

    Zhou Wang, Alan Bovik, Hamid Sheikh, and Eero Simon- celli. Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004. 7

    work page 2004

  35. [35]

    Invertible image signal processing

    Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. InCVPR, 2021. 2, 3, 8, 12

    work page 2021

  36. [36]

    High quality image reconstruction from RAW and JPEG image pair

    Lu Yuan and Jian Sun. High quality image reconstruction from RAW and JPEG image pair. InICCV, 2011. 2

    work page 2011

  37. [37]

    CycleISP: Real image restoration via improved data synthesis

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. CycleISP: Real image restoration via improved data synthesis. InCVPR, 2020. 2, 3, 8, 12

    work page 2020

  38. [38]

    Hui Zeng, Jianrui Cai, Lida Li, Zisheng Cao, and Lei Zhang. Learning image-adaptive 3D lookup tables for high perfor- mance photo enhancement in real-time.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2058– 2073, 2020. 5 14

    work page 2058