Pith. sign in

REVIEW 1 major objections 7 minor 60 references

AI Makes Plausible Images But Gets the Imaging Physics Wrong

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-09 17:40 UTC pith:IHGE4NRX

load-bearing objection First systematic benchmark showing frontier VLMs lag specialized methods on physics-grounded computational imaging, but the central comparative claim is overgeneralized relative to the baselines actually provided. the 1 major comments →

arxiv 2607.07189 v1 pith:IHGE4NRX submitted 2026-07-08 cs.AI

Does AI Understand Imaging? A Systematic Benchmark of Agentic AI for Computational Imaging Tasks

Ethan Chung , Chuanjun Zheng , Jasper Tan , Jingxi Li , Haopeng Zhang , Huaijin Chen This is my paper
classification cs.AI
keywords computational imagingvision-language modelsinverse problemsagentic AIbenchmarkforward operatorimage reconstructioncomputational sensing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces ImagingBench, a benchmark of 20 computational imaging tasks spanning ray and wave optics, image signal processing, inverse reconstruction, computational sensing, and calibration. The authors test three frontier multimodal systems (Gemini, GPT, Qwen) under three protocols: Expert (fixed expert-guided inverse reconstruction), Planner (adaptive per-image planning before execution), and Forward (forward-system simulation for consistency checking). The central finding is that these models exhibit a substantial gap between semantic visual competence and physically grounded imaging performance. On conventional image processing tasks like denoising and demosaicking, the models recover coarse scene structure but still trail specialized methods. On computational sensing tasks that require explicit inversion of a structured forward operator — lensless imaging, event-based reconstruction, time-of-flight depth imaging, and holography — the models fail badly, often producing outputs that look natural (good no-reference NIQE scores) while having poor reference-based fidelity (low PSNR and high LPIPS). Planner-guided prompting, where a reasoning model generates a tailored restoration instruction before execution, provides only modest and inconsistent gains over a fixed expert prompt, with near-zero average score deltas across model families. The paper argues that the main limitation is not prompting strategy but a lack of low-level imaging competence — the models rely on generic image priors rather than understanding how images are formed and how to invert that process.

Core claim

Frontier multimodal AI systems can generate visually plausible images but cannot reliably solve computational imaging problems that require understanding and inverting the physics of image formation. The gap is largest for computational sensing tasks where a structured forward operator must be explicitly inverted, and adaptive planning does not meaningfully close it. Visually natural outputs and physically correct reconstructions are not the same thing, and current models produce the former without achieving the latter.

What carries the argument

The benchmark formalizes each imaging task as a forward model x = A(z; m) + n, where z is the latent clean signal, A is the task-dependent forward operator, and n is noise. Three evaluation protocols isolate different competences: Expert tests fixed-prompt inverse execution, Planner tests adaptive per-image planning by a VLM before execution, and Forward tests whether the model can simulate the forward degradation process itself. Performance is measured with reference-based metrics (PSNR, SSIM, LPIPS) and the no-reference metric NIQE, then normalized into a unified score for cross-task comparison.

Load-bearing premise

The benchmark assumes that the three evaluation protocols and their fixed prompt templates adequately probe whether models understand imaging physics, rather than merely testing their sensitivity to how a task is described. The paper itself acknowledges that the protocols do not prove the model causally uses the forward model rather than generic image priors, yet the central conclusion about a gap between semantic and physical competence depends on this assumption holding.

What would settle it

If a frontier model were shown to produce high-fidelity reconstructions on the computational sensing tasks — lensless imaging, event-based reconstruction, time-of-flight, holography — with reference-based metrics matching specialized baselines, the claimed gap between visual plausibility and physical fidelity would collapse for that model class.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the gap between visual plausibility and physical fidelity is real, then benchmarks that score only perceptual quality will systematically overestimate AI competence on safety-critical imaging tasks in medicine, remote sensing, and scientific imaging.
  • The near-zero benefit of adaptive planning suggests that the bottleneck for agentic imaging is not instruction quality but the executor's lack of operator-aware reasoning, meaning progress will likely require architectural or training changes rather than better prompting.
  • The Forward protocol's results imply that models struggle not only to invert physics but to simulate it, which would undermine any pipeline that relies on a VLM to generate training data or consistency checks for imaging systems.
  • If specialized non-agentic methods remain substantially stronger, there is a practical case for hybrid systems where VLMs handle semantic reasoning and specialized solvers handle physical inversion, rather than end-to-end agentic pipelines.

Where Pith is reading between the lines

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

  • The finding that models produce natural-looking but physically incorrect outputs on sensing tasks suggests they are applying learned image priors rather than reasoning about the measurement process — a failure mode that is invisible to human evaluators and to no-reference quality metrics.
  • The fact that planner guidance rarely helps could indicate that the planner models themselves cannot diagnose the degradation type from visual inspection alone, which would mean the observe-plan-execute loop breaks at the observation stage rather than the execution stage.
  • A natural next test would be to give the planner explicit access to the forward operator parameters (sampling ratio, noise statistics, PSF) rather than asking it to infer them from the image, which would isolate whether the failure is in diagnosis or in execution.
  • The consistent superiority of specialized methods on sensing tasks raises the question of whether the gap would narrow if VLMs were fine-tuned on paired (measurement, reconstruction) data from specific forward operators, or whether the architectural mismatch between autoregressive/diffusion generation and inverse problem structure is more fundamental.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 7 minor

Summary. The paper introduces ImagingBench, a benchmark of 20 computational imaging tasks across five categories (ray and wave optics, image signal processing, inverse reconstruction, computational sensing, and calibration). Three evaluation protocols are defined: Expert (fixed expert-guided inverse reconstruction), Planner (planner-guided inverse reconstruction), and Forward (forward-system simulation for consistency checking). The authors evaluate Gemini, GPT, and Qwen image-editing models, finding that agentic models remain consistently weaker than specialized non-agentic baselines, especially on computational sensing tasks, and that planner guidance provides only modest gains over fixed expert prompts. The benchmark is well-motivated: it addresses a genuine gap between semantic vision benchmarks and physics-grounded computational imaging evaluation.

Significance. The paper makes a timely and useful contribution by providing the first unified benchmark spanning physical image formation, inverse reconstruction, calibration, and planner-executor evaluation for agentic AI. The three-protocol design (Expert, Planner, Forward) is a thoughtful framework for disentangling execution ability, per-instance planning, and forward-model consistency. The inclusion of ablation studies (noise sweeps, sampling-ratio sweeps, spherical-aberration sweeps) and cost analysis adds practical value. The finding that visually plausible outputs do not correspond to physically accurate solutions is important for the community. The benchmark infrastructure and task diversity are commendable.

major comments (1)
  1. The central claim that agentic models are 'consistently weaker than specialized methods' is directly supported by head-to-head comparison on only 9 of 20 tasks; 11 tasks in Table 3 report '—' for the Non-agent SoTA column. Critically, for computational sensing—the category the abstract emphasizes as where agentic models are 'especially' weaker—only lensless imaging (1/5 tasks) has a direct baseline. The other four sensing tasks (lightfield extrapolation, lightfield interpolation, event-based intensity, ToF depth) lack baselines entirely. The abstract specifically names 'event-based reconstruction' and 'time-of-flight imaging' as examples of where agentic models struggle most, yet these are exactly the tasks without specialized comparison. The claim therefore rests on low absolute PSNR values (e.g., ToF at ~5 dB, event-based at ~7-10 dB) rather than direct head-to-head evaluation. While 5
minor comments (7)
  1. Figure 1 caption refers to 'CIBench' but the benchmark is named 'ImagingBench' throughout the rest of the paper. This inconsistency should be corrected.
  2. Table 2 lists Comp. Gen. Holography under both 'Ray and Wave Optics' and 'Computational Sensing' categories. The rationale for this dual placement should be clarified.
  3. The normalized aggregate score (Section 3.3) uses hand-tuned weights (w_psnr=0.3, w_ssim=0.3, w_lpips=0.3, w_niqe=0.1) and normalization ranges (PSNR [15,40], NIQE [3,20]). A sensitivity analysis showing how the leaderboard ranking changes under alternative weight choices would strengthen confidence in the aggregate score's robustness.
  4. Section 3.4 mentions 'Nano Banana 2' as the Gemini model name but Table 3 and other references use 'Gemini-3.1-Flash-Image'. The naming should be made consistent.
  5. Table 3 uses color coding for metric values but the color scale is not always legible in print. Adding explicit numerical thresholds for 'good' and 'bad' ranges in the caption would aid interpretation.
  6. The paper states (Section 3.2) that the protocols 'do not by themselves prove that the model causally uses the provided forward model rather than generic image priors.' This is an important caveat that should be more prominently discussed, as the central conclusion about the gap between semantic and physical competence depends on this assumption.
  7. Section 6.5 mentions safety filter refusals from hosted models but does not quantify the refusal rate. Reporting the percentage of calls that triggered safety filters would be useful for practitioners.

Circularity Check

0 steps flagged

No circularity present.

full rationale

ImagingBench is an empirical benchmark paper, not a derivation paper. Its central claims rest on comparing agentic model outputs against external ground-truth data and independent task-specific baselines (e.g., FFDNet, ESRGAN, AutoLens). The forward models used to generate synthetic degradations (Poisson-Gaussian noise, Zernike aberrations, Bayer mosaicking, compressive sensing masks) are standard physics-based operators from the computational imaging literature, not fitted parameters repackaged as predictions. The normalized scoring (Eqs. 2-5) introduces hand-chosen clipping ranges and weights, but these are transparent aggregation choices for cross-task summarization, not fitted-to-data quantities presented as derived results. The paper explicitly states its protocols 'do not by themselves prove that the model causally uses the provided forward model rather than generic image priors,' which is a self-acknowledged limitation rather than a circular claim. No step in the benchmark construction or evaluation reduces to its own inputs by definition. The absence of non-agentic baselines on some tasks is a coverage gap (correctness risk), not a circularity issue.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The paper introduces no new physical entities or postulated objects. The free parameters are all in the scoring/normalization layer and do not affect the raw metric comparisons that support the central claim. The axioms are standard benchmark-design assumptions.

free parameters (4)
  • PSNR normalization range [15, 40] = 15 to 40 dB
    Used in Eq. 2 to clip and normalize PSNR scores; the range is chosen by hand without justification against a physical or perceptual threshold.
  • NIQE normalization range [3, 20] = 3 to 20
    Used in Eq. 5 to normalize NIQE; the range is hand-chosen.
  • Metric weights (w_psnr, w_ssim, w_lpips, w_niqe) = 0.3, 0.3, 0.3, 0.1
    Hand-assigned weights for combining normalized metrics into a single score; stated as 'base weights' without sensitivity analysis.
  • Strehl normalization threshold = 0.8
    Used in Norm_t = clip(Strehl/0.8, 0, 1) for lens design scoring; the 0.8 diffraction-limited threshold is a standard optics reference but the normalization mapping is a design choice.
axioms (3)
  • domain assumption Vision-language models that perform well on semantic visual tasks should be tested on physics-grounded tasks to assess transfer.
    The entire benchmark is motivated by this assumption (Section 1). It is reasonable but unproven that semantic competence should transfer to inverse problems.
  • ad hoc to paper The three protocols (Expert, Planner, Forward) are sufficient to disentangle execution ability, planning ability, and forward-model consistency.
    Section 3.2 defines these three settings as complementary, but the paper itself acknowledges they 'do not by themselves prove that the model causally uses the provided forward model.'
  • domain assumption Standardized 1024x1024 input resolution preserves task-defining structure for all subtasks.
    Section 3.4 states all inputs are standardized to 1024x1024. For tasks like lensless imaging or lightfield, this resolution may lose critical spatial information, though padding is used for some tasks.

pith-pipeline@v1.1.0-glm · 27369 in / 2526 out tokens · 353003 ms · 2026-07-09T17:40:27.425693+00:00 · methodology

0 comments
read the original abstract

Vision-language models (VLMs) and agentic AI have shown strong performance on semantic visual tasks, but it remains unclear whether they can handle the physics and inverse problems that underlie computational imaging. We present ImagingBench, a benchmark of 20 computational imaging tasks spanning five categories: ray and wave optics, image signal processing, inverse reconstruction, computational sensing, and calibration. ImagingBench evaluates three complementary settings: Expert, fixed expert-guided inverse reconstruction; Planner, planner-guided inverse reconstruction; and Forward, forward-system simulation for consistency checking. We benchmark leading proprietary and open-source image-centric multimodal systems, including Gemini, GPT, and Qwen, and compare them with representative task-specific non-agentic baselines. Across tasks, agentic models remain consistently weaker than specialized methods, especially on computational sensing problems such as lensless imaging, event-based reconstruction, time-of-flight imaging, and holography. Planner guidance provides only modest and inconsistent gains over the fixed-prompt Expert baseline. Although the models often generate visually plausible outputs, their reference-based fidelity remains poor, revealing a substantial gap between semantic visual competence and physically grounded imaging performance. ImagingBench provides a unified testbed for measuring this gap and tracking progress in agentic AI for computational imaging.

Figures

Figures reproduced from arXiv: 2607.07189 by Chuanjun Zheng, Ethan Chung, HaoPeng Zhang, Huaijin Chen, Jasper Tan, Jingxi Li.

Figure 1
Figure 1. Figure 1: Overview of ImagingBench: Our benchmark comprises 20 subtasks spanning five computational imaging categories. Each panel shows representative tasks in the corresponding category, illustrating how the benchmark spans the computational imaging pipeline from scene acquisition and optical encoding to sensing, image processing, reconstruction, and calibration. lines, and analyze where foundation-model compe￾ten… view at source ↗
Figure 2
Figure 2. Figure 2: Representative qualitative examples from ImagingBench: For each example, we show the task prompt and representative outputs produced by three frontier agentic AI systems (GPT, Gemini, and Qwen), highlighting the breadth of ImagingBench across physically grounded inverse problems and structured imaging-reasoning tasks. The examples also illustrate substantial variation in output fidelity, physical plausibil… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the ImagingBench evaluation pipeline: Each input task is paired with a selected dataset and prompt template, then evaluated under three protocols: the Expert protocol (expert-guided inverse reconstruction), the Planner protocol (planner-guided inverse reconstruction), and the Forward protocol (forward-system simulation). In Planner and Forward, a frontier agentic AI planner generates image-spec… view at source ↗
Figure 4
Figure 4. Figure 4: First, we adopt established datasets when they match [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Semantic scene distribution across ImagingBench subtasks, shown as [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Unified-score radar comparison across representative image-reconstruction subtasks. Each axis reports task-level unified score in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-task distributions of the unified score in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of representative ray-optics lens designs under different [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual comparison of outputs from GPT, Gemini, and Qwen on three representative ImagingBench tasks: compressive sensing (top), demosaicking [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Representative CGH failure case. The agents recover coarse structure, [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

60 extracted references · 60 canonical work pages · 6 internal anchors

  1. [1]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei, “Imagenet large scale visual recognition challenge,” International Journal of Computer Vision, vol. 115, no. 3, pp. 211–252, 2015

    work page 2015

  2. [2]

    The PASCAL visual object classes (VOC) challenge,

    M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zis- serman, “The PASCAL visual object classes (VOC) challenge,” International Journal of Computer Vision, vol. 88, no. 2, pp. 303–338, 2010

    work page 2010

  3. [3]

    Microsoft COCO: Common objects in context,

    T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P . Perona, D. Ramanan, P . Doll´ar, and C. L. Zitnick, “Microsoft COCO: Common objects in context,” inEuropean Conference on Computer Vision (ECCV), 2014, pp. 740–755

    work page 2014

  4. [4]

    GLUE: A multi-task benchmark and analysis platform for nat- ural language understanding,

    A. Wang, A. Singh, J. Michael, F. Hill, O. Levy, and S. R. Bowman, “GLUE: A multi-task benchmark and analysis platform for nat- ural language understanding,” inProceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, 2018, pp. 353–355

    work page 2018

  5. [5]

    Beyond the imitation game: Quantifying and extrapolating the capabilities of language models,

    A. Srivastava, A. Rastogi, A. Raoet al., “Beyond the imitation game: Quantifying and extrapolating the capabilities of language models,”Transactions on Machine Learning Research, 2023. [Online]. Available: https://openreview.net/forum?id=uyTL5Bvosj

    work page 2023

  6. [6]

    MMBench: Is your multi-modal model an all-around player?

    Y. Liu, H. Duan, Y. Zhang, B. Li, S. Zhang, W. Zhao, Y. Yuan, J. Wang, C. He, Z. Liu, K. Chen, and D. Lin, “MMBench: Is your multi-modal model an all-around player?” inEuropean Conference on Computer Vision (ECCV), 2024, pp. 216–233

    work page 2024

  7. [7]

    MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for expert AGI,

    X. Yue, Y. Ni, K. Zhang, T. Zheng, R. Liu, G. Zhang, S. Stevens, D. Jiang, W. Ren, Y. Sun, C. Wei, B. Yu, R. Yuan, R. Sun, M. Yin, B. Zheng, Z. Yang, Y. Liu, W. Huang, H. Sun, Y. Su, and W. Chen, “MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for expert AGI,” inProceedings of the IEEE/CVF Conference on Computer Vision a...

    work page 2024

  8. [8]

    AgentBench: Evaluating LLMs as Agents

    X. Liu, H. Yu, H. Zhang, Y. Xu, X. Lei, H. Lai, Y. Gu, H. Ding, K. Men, K. Yang, S. Zhang, X. Deng, A. Zeng, Z. Du, C. Zhang, S. Shen, T. Zhang, Y. Su, H. Sun, M. Huang, Y. Dong, and J. Tang, “AgentBench: Evaluating LLMs as agents,” 2023. [Online]. Available: https://arxiv.org/abs/2308.03688

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

    GAIA: A benchmark for general AI assistants,

    G. Mialon, C. Fourrier, T. Wolf, Y. LeCun, and T. Scialom, “GAIA: A benchmark for general AI assistants,” inInternational Conference on Learning Representations (ICLR), 2024. [Online]. Available: https://openreview.net/forum?id=fibxvahvs3

    work page 2024

  10. [10]

    PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding

    W. Chow, J. Mao, B. Li, D. Seita, V . Guizilini, and Y. Wang, “Physbench: Benchmarking and enhancing vision-language models for physical world understanding,” inICLR, 2025. [Online]. Available: https://arxiv.org/abs/2501.16411

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  11. [11]

    MMMG: a Comprehensive and Reliable Evaluation Suite for Multitask Multimodal Generation

    J. Yao, Y. Hu, Y. Yi, B. Han, S. Feng, G. Yang, B. Wen, R. Krishna, L. L. Wang, Y. Tsvetkov, N. A. Smith, and B. Zhu, “Mmmg: a comprehensive and reliable evaluation suite for multitask multimodal generation,” 2025. [Online]. Available: https://arxiv.org/abs/2505.17613

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    Com- putational imaging and artificial intelligence: The next revolution of mobile vision,

    J. Suo, W. Zhang, J. Gong, X. Yuan, D. J. Brady, and Q. Dai, “Com- putational imaging and artificial intelligence: The next revolution of mobile vision,”Proceedings of the IEEE, vol. 111, no. 12, pp. 1607– 1639, 2023

    work page 2023

  13. [13]

    do: A differentiable engine for deep lens design of computational imaging systems,

    C. Wang, N. Chen, and W. Heidrich, “do: A differentiable engine for deep lens design of computational imaging systems,”IEEE Transactions on Computational Imaging, vol. 8, pp. 905–916, 2022

    work page 2022

  14. [14]

    Curriculum learning for ab initio deep learned refractive optics,

    X. Yang, Q. Fu, and W. Heidrich, “Curriculum learning for ab initio deep learned refractive optics,”Nature communications, vol. 15, no. 1, p. 6572, 2024

    work page 2024

  15. [15]

    Bayesian-based iterative method of image restoration,

    W. H. Richardson, “Bayesian-based iterative method of image restoration,”Journal of the Optical Society of America, vol. 62, no. 1, pp. 55–59, 1972

    work page 1972

  16. [16]

    An iterative technique for the rectification of observed distributions,

    L. B. Lucy, “An iterative technique for the rectification of observed distributions,”The Astronomical Journal, vol. 79, pp. 745–754, 1974

    work page 1974

  17. [17]

    Nonlinear total variation based noise removal algorithms,

    L. I. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based noise removal algorithms,”Physica D: Nonlinear Phenomena, vol. 60, no. 1-4, pp. 259–268, 1992

    work page 1992

  18. [18]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,

    K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,”IEEE transactions on image processing, vol. 26, no. 7, pp. 3142–3155, 2017

    work page 2017

  19. [19]

    Learning a single convolutional super-resolution network for multiple degradations,

    K. Zhang, W. Zuo, and L. Zhang, “Learning a single convolutional super-resolution network for multiple degradations,” inProceed- ings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3262–3271

    work page 2018

  20. [20]

    VQA: Visual question answering,

    S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, and D. Parikh, “VQA: Visual question answering,” inProceedings of the IEEE International Conference on Computer Vision (ICCV), 2015, pp. 2425–2433

    work page 2015

  21. [21]

    nocaps: Novel object captioning at scale,

    H. Agrawal, K. Desai, Y. Wang, X. Chen, R. Jain, M. Johnson, D. Batra, D. Parikh, S. Lee, and P . Anderson, “nocaps: Novel object captioning at scale,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 8948–8957

    work page 2019

  22. [22]

    MedMNIST v2: A large-scale lightweight benchmark for 2d and 3d biomedical image classification,

    J. Yang, R. Shi, D. Wei, Z. Liu, L. Zhao, B. Ke, H. Pfister, and B. Ni, “MedMNIST v2: A large-scale lightweight benchmark for 2d and 3d biomedical image classification,”Scientific Data, vol. 10, no. 1, p. 41, 2023

    work page 2023

  23. [23]

    CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison,

    J. Irvin, P . Rajpurkar, M. Ko, Y. Yu, S. Ciurea-Ilcus, C. Chute, H. Marklund, B. Haghgoo, R. Ball, K. Shpanskaya, J. Seekins, D. A. Mong, S. S. Halabi, J. K. Sandberg, R. Jones, D. B. Larson, C. P . Langlotz, B. N. Patel, M. P . Lungren, and A. Y. Ng, “CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison,” inProceeding...

    work page 2019

  24. [24]

    MIMIC- CXR, a de-identified publicly available database of chest radio- graphs with free-text reports,

    A. E. W. Johnson, T. J. Pollard, S. J. Berkowitz, N. R. Greenbaum, M. P . Lungren, C.-y. Deng, R. G. Mark, and S. Horng, “MIMIC- CXR, a de-identified publicly available database of chest radio- graphs with free-text reports,”Scientific Data, vol. 6, p. 317, 2019

    work page 2019

  25. [25]

    VinDr-CXR: 14 An open dataset of chest x-rays with radiologist’s annotations,

    H. Q. Nguyen, K. Lam, L. T. Le, H. H. Nguyen, H. H. Pham, H. Tong, D. Dinh, D. Nguyen, M. Dao, V . Vuet al., “VinDr-CXR: 14 An open dataset of chest x-rays with radiologist’s annotations,” Scientific Data, vol. 9, no. 1, p. 429, 2022

    work page 2022

  26. [26]

    The mul- timodal brain tumor image segmentation benchmark (BRATS),

    B. H. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, J. Kirby, Y. Burren, N. Porz, J. Slotboom, R. Wiestet al., “The mul- timodal brain tumor image segmentation benchmark (BRATS),” IEEE Transactions on Medical Imaging, vol. 34, no. 10, pp. 1993–2024, 2015

    work page 1993

  27. [27]

    A dataset of clinically generated visual questions and answers about radiology images,

    J. J. Lau, S. Gayen, A. Ben Abacha, and D. Demner-Fushman, “A dataset of clinically generated visual questions and answers about radiology images,”Scientific Data, vol. 5, p. 180251, 2018

    work page 2018

  28. [28]

    SLAKE: A semantically-labeled knowledge-enhanced dataset for medical visual question answering,

    B. Liu, L.-M. Zhan, L. Xu, L. Ma, Y. Yang, and X.-M. Wu, “SLAKE: A semantically-labeled knowledge-enhanced dataset for medical visual question answering,” in2021 IEEE 18th International Sympo- sium on Biomedical Imaging (ISBI), 2021, pp. 1650–1654

    work page 2021

  29. [29]

    PathVQA: 30000+ Questions for Medical Visual Question Answering

    X. He, Y. Zhang, L. Mou, E. Xing, and P . Xie, “PathVQA: 30000+ questions for medical visual question answering,” 2020. [Online]. Available: https://arxiv.org/abs/2003.10286

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  30. [30]

    PMC-VQA: Visual Instruction Tuning for Medical Visual Question Answering

    X. Zhang, C. Wu, Z. Zhao, W. Lin, Y. Zhang, Y. Wang, and W. Xie, “PMC-VQA: Visual instruction tuning for medical visual question answering,” 2023. [Online]. Available: https://arxiv.org/abs/2305.10415

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  31. [31]

    GMAI-MMBench: A comprehensive multimodal evaluation benchmark towards general medical ai,

    P . Chen, J. Ye, G. Wang, Y. Li, Z. Deng, W. Li, T. Li, H. Duan, Z. Huang, Y. Su, B. Wang, S. Zhang, B. Fu, J. Cai, B. Zhuang, E. J. Seibel, Y. Qiao, and J. He, “GMAI-MMBench: A comprehensive multimodal evaluation benchmark towards general medical ai,” inAdvances in Neural Information Processing Systems, vol. 37, 2024. [Online]. Available: https://proceed...

    work page 2024

  32. [32]

    MMMG: A massive, multidisciplinary, multi-tier generation benchmark for text-to-image reasoning,

    Y. Luo, Y. Yuan, J. Chen, H. Cai, Z. Yue, Y. Yang, F. Z. Daha, J. Li, and Z. Lian, “MMMG: A massive, multidisciplinary, multi-tier generation benchmark for text-to-image reasoning,”

    work page

  33. [33]

    Available: https://arxiv.org/abs/2506.10963

    [Online]. Available: https://arxiv.org/abs/2506.10963

    work page arXiv

  34. [34]

    Hybrid optical-electronic convolutional neural networks with optimized diffractive optics for image classification,

    J. Chang, V . Sitzmann, X. Dun, W. Heidrich, and G. Wetzstein, “Hybrid optical-electronic convolutional neural networks with optimized diffractive optics for image classification,”Scientific reports, vol. 8, no. 1, p. 12324, 2018

    work page 2018

  35. [35]

    Vision-language model guided image restoration,

    C. Yang, R. Dong, and K.-M. Lam, “Vision-language model guided image restoration,” 2025. [Online]. Available: https: //arxiv.org/abs/2512.17292

    work page arXiv 2025

  36. [36]

    Optiagent: A physics-driven agentic framework for automated optical design,

    Y. Geng, L. Sun, Y. Gao, X. Hu, Z. Yi, X. Qian, W. Hu, J. Bai, and K. Wang, “Optiagent: A physics-driven agentic framework for automated optical design,” 2026. [Online]. Available: https://arxiv.org/abs/2602.23761

    work page arXiv 2026

  37. [37]

    Towards real- time photorealistic 3d holography with deep neural networks,

    L. Shi, B. Li, C. Kim, P . Kellnhofer, and W. Matusik, “Towards real- time photorealistic 3d holography with deep neural networks,” Nature, vol. 591, no. 7849, pp. 234–239, 2021

    work page 2021

  38. [38]

    Uni- fied reconstruction of static and dynamic scenes from events,

    Q. Gao, P . Duan, H. Lou, M. Teng, Z. Cai, X. Chen, and B. Shi, “Uni- fied reconstruction of static and dynamic scenes from events,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 27 914–27 923

    work page 2025

  39. [39]

    Depth restoration in under-display time-of-flight imaging,

    X. Qiao, C. Ge, P . Deng, H. Wei, M. Poggi, and S. Mattoccia, “Depth restoration in under-display time-of-flight imaging,”IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 5, pp. 5668–5683, 2022

    work page 2022

  40. [40]

    Learned reconstructions for practical mask-based lens- less imaging,

    K. Monakhova, J. Yurtsever, G. Kuo, N. Antipa, K. Yanny, and L. Waller, “Learned reconstructions for practical mask-based lens- less imaging,”Optics express, vol. 27, no. 20, pp. 28 075–28 090, 2019

    work page 2019

  41. [41]

    When color constancy goes wrong: Correcting improperly white-balanced im- ages,

    M. Afifi, B. Price, S. Cohen, and M. S. Brown, “When color constancy goes wrong: Correcting improperly white-balanced im- ages,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 1535–1544

    work page 2019

  42. [42]

    Beyond joint demosaick- ing and denoising: An image processing pipeline for a pixel-bin image sensor,

    S. A Sharif, R. A. Naqvi, and M. Biswas, “Beyond joint demosaick- ing and denoising: An image processing pipeline for a pixel-bin image sensor,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 233–242

    work page 2021

  43. [43]

    Burst photography for high dynamic range and low-light imaging on mobile cameras,

    S. W. Hasinoff, D. Sharlet, R. Geiss, A. Adams, J. T. Barron, F. Kainz, J. Chen, and M. Levoy, “Burst photography for high dynamic range and low-light imaging on mobile cameras,”ACM Transactions on Graphics (Proc. SIGGRAPH Asia), vol. 35, no. 6, 2016

    work page 2016

  44. [44]

    Fpa-cs: Focal plane array-based compressive imaging in short-wave infrared,

    H. Chen, M. Salman Asif, A. C. Sankaranarayanan, and A. Veer- araghavan, “Fpa-cs: Focal plane array-based compressive imaging in short-wave infrared,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 2358–2366

    work page 2015

  45. [45]

    Resolution-robust large mask inpainting with fourier convolutions,

    R. Suvorov, E. Logacheva, A. Mashikhin, A. Remizova, A. Ashukha, A. Silvestrov, N. Kong, H. Goka, K. Park, and V . Lem- pitsky, “Resolution-robust large mask inpainting with fourier convolutions,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2022, pp. 2149–2159

    work page 2022

  46. [46]

    Gemini 3.1 flash-image model card,

    Google DeepMind, “Gemini 3.1 flash-image model card,” https://deepmind.google/models/model-cards/ gemini-3-1-flash-image/, 2026, accessed: 2026-02-26

    work page 2026

  47. [47]

    [Online]

    OpenAI,GPT-Image 1.5 Model Documentation, OpenAI, San Fran- cisco, CA, 2026, accessed: 2026-02-26. [Online]. Available: https: //developers.openai.com/api/docs/models/gpt-image-1.5

    work page 2026

  48. [48]

    Qwen-Image Technical Report

    C. Wu, J. Li, J. Zhou, J. Lin, K. Gao, K. Yan, S. ming Yin, S. Bai, X. Xu, Y. Chen, Y. Chen, Z. Tang, Z. Zhang, Z. Wang, A. Yang, B. Yu, C. Cheng, D. Liu, D. Li, H. Zhang, H. Meng, H. Wei, J. Ni, K. Chen, K. Cao, L. Peng, L. Qu, M. Wu, P . Wang, S. Yu, T. Wen, W. Feng, X. Xu, Y. Wang, Y. Zhang, Y. Zhu, Y. Wu, Y. Cai, and Z. Liu, “Qwen-image technical repo...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  49. [49]

    A high-quality de- noising dataset for smartphone cameras,

    A. Abdelhamed, S. Lin, and M. S. Brown, “A high-quality de- noising dataset for smartphone cameras,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    work page 2018

  50. [50]

    Image demosaicing: A system- atic survey,

    X. Li, B. Gunturk, and L. Zhang, “Image demosaicing: A system- atic survey,” inVisual Communications and Image Processing 2008, vol. 6822. SPIE, 2008, pp. 489–503

    work page 2008

  51. [51]

    Deep multi-scale convolutional neural network for dynamic scene deblurring,

    S. Nah, T. H. Kim, and K. M. Lee, “Deep multi-scale convolutional neural network for dynamic scene deblurring,” inCVPR, July 2017

    work page 2017

  52. [52]

    The stanford light field archive (2016),

    Stanford Computer Graphics Laboratory, “The stanford light field archive (2016),” https://lightfields.stanford.edu/LF2016. html, 2016, accessed: 2026-02-26

    work page 2016

  53. [53]

    Ntire 2017 challenge on single image super-resolution: Dataset and study,

    E. Agustsson and R. Timofte, “Ntire 2017 challenge on single image super-resolution: Dataset and study,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017

    work page 2017

  54. [54]

    Learning photo- graphic global tonal adjustment with a database of input / output image pairs,

    V . Bychkovsky, S. Paris, E. Chan, and F. Durand, “Learning photo- graphic global tonal adjustment with a database of input / output image pairs,” inThe Twenty-Fourth IEEE Conference on Computer Vision and Pattern Recognition, 2011

    work page 2011

  55. [55]

    Phasecam3d—learning phase masks for pas- sive single view depth estimation,

    Y. Wu, V . Boominathan, H. Chen, A. Sankaranarayanan, and A. Veeraraghavan, “Phasecam3d—learning phase masks for pas- sive single view depth estimation,” in2019 IEEE International Conference on Computational Photography (ICCP). IEEE, 2019, pp. 1–12

    work page 2019

  56. [56]

    A physics-based noise formation model for extreme low-light raw denoising,

    K. Wei, Y. Fu, J. Yang, and H. Huang, “A physics-based noise formation model for extreme low-light raw denoising,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 2758–2767

    work page 2020

  57. [57]

    Deep white-balance editing,

    M. Afifi and M. S. Brown, “Deep white-balance editing,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020

    work page 2020

  58. [58]

    Ffdnet: Toward a fast and flexible solution for cnn-based image denoising,

    K. Zhang, W. Zuo, and L. Zhang, “Ffdnet: Toward a fast and flexible solution for cnn-based image denoising,”IEEE Transactions on Image Processing, vol. 27, no. 9, pp. 4608–4622, 2018

    work page 2018

  59. [59]

    Banet: A blur-aware attention network for dynamic scene deblurring,

    F.-J. Tsai, Y.-T. Peng, C.-C. Tsai, Y.-Y. Lin, and C.-W. Lin, “Banet: A blur-aware attention network for dynamic scene deblurring,” IEEE Transactions on Image Processing, vol. 31, pp. 6789–6799, 2022

    work page 2022

  60. [60]

    Esrgan: Enhanced super-resolution generative adversarial networks,

    X. Wang, K. Yu, S. Wu, J. Gu, Y. Liu, C. Dong, Y. Qiao, and C. Change Loy, “Esrgan: Enhanced super-resolution generative adversarial networks,” inEuropean Conference on Computer Vision (ECCV), 2018, pp. 0–0

    work page 2018