SegDem: Segmentation helps Demosaicing

Haijin Zeng; Jie Liu; Jiezhang Cao; Jingyong Su; Kai Zhang; Ping Chen; Xiangming Wang; Yongyong Chen

REVIEW 4 major objections 5 minor 61 references

SegDem: Segmentation helps Demosaicing

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read SegDem claims that transferring region- and boundary-aware decoder representations from instance segmentation, anchored by frozen DINOv2 features, consistently improves demosaicing across three backbones and two CFA layouts.

desk verdict SegDem is a genuinely new combination of segmentation pretraining and DINOv2 anchoring for demosaicing that shows consistent but modest gains; the central causal claim needs a matched pretraining control before it is fully convincing. read the letter →

arxiv 2608.07916 v1 pith:73EPEB7Y submitted 2026-08-08 cs.CV

Ping Chen , Xiangming Wang , Yongyong Chen , Jiezhang Cao , Kai Zhang , Jingyong Su , Jie Liu , Haijin Zeng This is my paper · ORCID

classification cs.CV

keywords imagedemosaicinginstancesegmentationrepresentationtransferdecoderpretrainingDINOv2colorfilterarrayQuadBayercross-tasklearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that high-level visual understanding can improve low-level image reconstruction. The authors argue that segmentation and demosaicing are complementary views of the same scene structure, so the region- and boundary-aware representations learned by an instance segmentation decoder should transfer to demosaicing. They propose SegDem, a two-stage framework: first pretrain a RAW-conditioned decoder with instance masks, then fine-tune it for demosaicing while anchoring both task features to a frozen DINOv2 representation space. The claim is that this transfer consistently improves demosaicing quality on both Single- and Quad-Bayer color filter arrays, across three different backbone families, without adding segmentation outputs at inference. If true, it shows a cheap route to better demosaicing: borrow structural knowledge from segmentation rather than generating or hallucinating details.

What carries the argument

The load-bearing mechanism is the two-stage cross-task decoder transfer. Stage 1 takes a RAW-conditioned encoder-decoder and trains it to predict RGB-coded instance masks (Gen2Seg-style) on COCO, forcing decoder features to organize into object regions and boundaries. Stage 2 initializes the demosaicing network with those pretrained decoder weights and fine-tunes on RAW/RGB pairs, while frozen DINOv2 tokens supervise both segmentation- and demosaicing-conditioned decoder features through a masked Smooth-L1 token loss. The RAW condition branch preserves sensor phase via pattern-aware CFA packing, pixel binning, and a sparse RAW correction module. The central idea is that segmentation- and demosaicing-conditioned features are never matched directly to each other; they are independently anchored to the same frozen DINOv2 space, preserving structural organization across the task transition without forcing the two feature distributions to coincide.

What would settle it

Train the same two-stage protocol on native RAW captures from several camera models with real sensor noise and evaluate against the same baselines; if segmentation-pretrained instantiations do not beat their from-scratch counterparts in sRGB PSNR or LPIPS, the claim that transferred structural representations improve demosaicing would be unsupported.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that instance segmentation supplies a transferable structural prior for demosaicing: region- and boundary-aware decoder representations learned from COCO instance masks carry over to RAW-conditioned reconstruction, and anchoring both task views to a shared frozen DINOv2 representation preserves that structure during fine-tuning. The paper reports consistent PSNR/SSIM/LPIPS gains over strong baselines on Single- and Quad-Bayer CFAs across three backbone families, with the largest improvements in rendered sRGB and LPIPS, which the authors interpret as evidence that the prior helps exactly where local color continuation is ambiguous—at boundaries, thin structures, and repetitive patterns.

Load-bearing premise

The synthetic RAW pipeline—a fixed inverse PTP transform with gamma 2.2, a fixed 3x3 color correction matrix, and tiny Gaussian sensor noise—is representative enough of real camera data that the measured gains will survive the transfer to practice; the authors note this limitation in Appendix G.

Editorial extensions

If this is right

Inference-time cost is unchanged: the segmentation head, DINOv2 encoder, and instruction encoder are removed after training, so the improvement comes free at deployment.
The transfer works across convolutional, Transformer, and state-space backbones, so it is not an artifact of one architecture family.
Both Single-Bayer and Quad-Bayer CFA layouts improve, with the largest relative gains in rendered sRGB and LPIPS, meaning the benefit shows up where color and boundary errors are visible.
Because the reconstruction path remains explicitly RAW-conditioned and the segmentation head is discarded, the framework does not add semantic hallucination at inference.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same two-stage recipe could be applied to other ambiguity-limited reconstructions (denoising, super-resolution, deblurring), since those tasks also suffer from boundary and repetitive-pattern ambiguity.
The paper's supplementary finding that demosaicing pretraining also helps segmentation suggests a bidirectional loop may be more powerful than the one-way transfer the authors test; a joint or alternating training objective could push both tasks further.
DINOv2 is only one possible structural anchor; other self-supervised spatial feature extractors might give similar or better anchoring, and the paper does not test alternatives.
The largest gains appear after sRGB rendering, so the practical impact likely depends on the downstream ISP; a camera pipeline whose rendering differs from the fixed PTP operator could dilute or amplify the benefit.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: SegDem is a well-put-together empirical paper with a genuinely new combination — instance-segmentation decoder pretraining, DINOv2-anchored cross-task alignment, and RAW-conditioned reconstruction — applied to demosaicing. The gains are real but modest, and the central causal claim is not fully nailed down. Worth a serious referee, but the authors should be asked for a matched pretraining control before the core message is accepted.

What's new: I haven't seen this exact cross-task decoder transfer for demosaicing. Pretraining the decoder with RGB-coded COCO masks, then discarding the segmentation head and fine-tuning on RAW/RGB pairs, is a clean instantiation of the 'understanding helps reconstruction' idea. The DINOv2 token alignment used only at training time is a sensible way to preserve structural organization without forcing a semantic output at inference. The paper also does the right things empirically: three different backbones (conv, transformer, SSM), two CFA layouts, external SANet-derived and MIT-Hard sets, and a component ablation in Table 5. The gains are not huge — up to ~1.4 dB in the most favorable case, sub-0.1 dB in the weakest — but they are consistent.

Where the soft spots are: the main one is Table 4. The 'Seg. pretrain: No' condition trains from scratch for 10 epochs on LSDIR, while 'Yes' gets an extra, unspecified COCO pretraining phase before the same 10-epoch fine-tune. So you are comparing more total gradient updates plus a warm start from a large natural-image corpus against a from-scratch run. That does not isolate the segmentation objective. A matched control — same COCO RAW data, same amount of pretraining, but a non-semantic objective (e.g., direct demosaicing or DINO token regression) — is needed to claim that region/boundary structure is what transfers. Without it, the gains in Table 4 could be generic transfer learning. Your point about MambaIR's linear PSNR delta of +0.08 dB and SSIM delta of +0.0000 is well taken; the effect is thin in the weakest case. Also, no code is released yet, no error bars or multi-seed runs, and the baseline retraining protocol (were the baselines retrained on the same split?) is not fully specified. The authors do admit the synthetic-data limitation in Appendix G, which is honest.

Is the central argument dead? No. The paper is a competent empirical contribution; the concern is attribution, not fabrication. A referee can ask for one additional experiment that would settle it.

Who this is for: people working on demosaicing or on transferring high-level representations to low-level reconstruction. It deserves a peer-review slot; with a matched pretraining control and code release, it could be a solid paper.

Recommendation: send to review, with the request for a controlled pretraining ablation.

Referee Report

4 major / 5 minor

Summary. The paper proposes SegDem, a two-stage framework for image demosaicing in which a shared decoder is first pretrained with instance-segmentation supervision on COCO-derived RAW data, then fine-tuned for RAW-conditioned linear-RGB reconstruction on LSDIR. A frozen DINOv2 encoder provides a common token-level representation space that both segmentation- and demosaicing-conditioned decoder features are aligned to at training time. SegDem is instantiated with three backbones (EDMUNet, Restormer, MambaIR) and evaluated on synthetic Single- and Quad-Bayer benchmarks, on SANet-derived external pairs, and on MIT-Hard. The main claim is that region- and boundary-aware decoder representations transferred from segmentation consistently improve demosaicing quality across architectures and CFA layouts.

Significance. If the causal claim holds, the paper offers a plausible and novel way to inject high-level structural priors into low-level reconstruction without requiring semantic outputs at inference, and the external evaluations on SANet and MIT-Hard are a genuine strength. The design is reproducible in principle: code will be released, the training protocol is described in appendix detail, and the paper includes component ablations that separate several contributions. The main weakness is that the central attribution of the gains to the segmentation objective itself is not yet supported by a controlled comparison, and the statistical evidence for 'consistent' gains is thin in several cells. These issues are fixable with additional experiments, so the result remains plausible rather than established.

major comments (4)

[§4.5, Table 4 and Appendix A] The central causal claim that instance-aware segmentation pretraining is what improves demosaicing is under-identified. The 'Yes' condition receives an additional COCO pretraining phase (whose epoch count is not given in Appendix A) followed by 10 epochs of fine-tuning, while the 'No' condition is trained from scratch on LSDIR for 10 epochs. The 'Yes' condition therefore has more total gradient updates and a warm start from a large natural-image corpus. A matched control is needed: pretrain on the same COCO RAW data with a non-semantic objective (e.g., demosaicing from the mosaic, or regressing DINOv2 tokens directly) and then apply the same Stage-2 fine-tuning. Without such a control, the gains in Table 4 could be generic transfer learning or extra optimization rather than region- and boundary-aware structural transfer.
[§4.5, Table 4] Several reported deltas are very small — for MambaIR on Quad Bayer, linear PSNR improves by +0.0826 dB and SSIM by +0.0000 — and no standard deviations or multiple-seed results are reported anywhere. The claim of consistent improvements across backbones and CFA layouts therefore lacks statistical support. Please report results over at least three seeds (or equivalent variance information) for the main ablations, and draw conclusions about consistency only for effects that exceed the noise level.
[§4.1 and Tables 1–3] The baseline comparison protocol is not specified. The paper does not state whether Jd3Net, BMTNet, FFTNet, and ms-demosaic were retrained on the same LSDIR 7:1:2 split with the same training schedule, or whether their numbers come from official checkpoints trained on other data. If the latter, the comparisons in Tables 1–3 are potentially confounded by different training data and schedule. Please clarify the protocol and, where official checkpoints are used, justify that the comparison is fair; ideally, retrain the baselines under the same protocol.
[§4.5, Table 5 and §3.4] The ablation in Table 5 shows that removing DINOv2 alignment from both stages causes a substantial drop, yet the paper does not fully separate the benefit of the segmentation pretraining from the benefit of the DINOv2 alignment when both are present. Specifically, a 'no segmentation pretraining, with DINOv2 alignment' condition is reported (the 'No' rows of Table 4), but it still uses the same Stage-2 DINO alignment; a complementary 'segmentation pretraining, no DINO alignment in either stage' condition would clarify whether the gains attributed to segmentation persist without the DINO anchor. This additional condition would sharpen the causal story.

minor comments (5)

[§1] The phrase 'state-space!' in the introduction contains a stray exclamation mark; it should read 'state-space'.
[§4.4] The sentence 'Generally, we use the MIT-Hard dataset' is awkward and likely intended as 'We additionally use the MIT-Hard dataset'; please rephrase.
[§4.4 and Table 3] The dataset is referred to as 'MIT-Hard' in the text and 'Moiré-Hard' in the same section; please standardize the name.
[Table 5] Several numeric entries in the table are missing spacing (e.g., '33.8543 0.92240.0903' and '32.96120.9203'), making the table hard to read; please fix the formatting.
[Appendix G] The limitation that training and evaluation rely on synthetic or re-mosaicked RAW observations is acknowledged, and this is good; however, the abstract and conclusion currently state improvements without this caveat, so consider adding a sentence in the conclusion that restates the synthetic-data limitation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SegDem's claims are empirical, with held-out and external evaluations; the Table 4 pretraining comparison is a control issue, not a definitional reduction.

full rationale

SegDem's derivation chain is empirical rather than analytic. Stage 1 learns decoder weights from COCO instance masks with Eq. 3; Stage 2 initializes the demosaicing network from those weights and optimizes reconstruction losses (Eq. 4) plus DINOv2 anchoring (Eqs. 5-6) on LSDIR RAW/RGB pairs. The reported quantities — linear and sRGB PSNR/SSIM and LPIPS on the held-out LSDIR test split, SANet-derived external pairs, and MIT-Hard — are computed against ground-truth linear RGB or fixed-PTP renders, not against any parameter fitted in SegDem. DINOv2 and the Jd3Net proxy are frozen, external tools used as supervision or input, not restatements of the output; the DINO alignment is a training-time regularizer, not a definition of demosaicing quality. The only near-concern is Table 4, where the 'No' condition is a 10-epoch from-scratch baseline while 'Yes' adds COCO pretraining and hence more total gradient updates; this is an experimental attribution and control issue, not a circular derivation, because the improvement is still measured on data disjoint from both training stages. Self-citations in the related work are contextual and none is load-bearing for the central transfer claim. No uniqueness theorem or ansatz is imported via self-citation, and no equation reduces to its own input.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method depends on several hand-chosen hyperparameters (loss weights, token mask ratio, tanh bound) and on domain assumptions about segmentation-to-demosaicing transfer and synthetic RAW realism. No new physical entities are introduced.

free parameters (7)

lambda_inst = 1.0
Instance loss weight in Stage 1, chosen by hand.
lambda_seg_repr = 0.1
Weight for DINOv2 alignment in the segmentation pretraining stage.
lambda_demo_repr = 0.1
Weight for DINOv2 alignment in the demosaicing fine-tuning stage.
lambda_char, lambda_srgb, lambda_lpips, lambda_edge = 1.0, 1.0, 0.2, 0.1
Weights of the four reconstruction losses in Eq. (4), chosen by hand.
DINO token mask ratio = 0.4
Fraction of token locations randomly sampled for DINO alignment; ablating it changes results (Table 5).
RAW correction tanh bound = 0.25
Bounds the residual predicted by the sparse RAW correction module.
fixed noise sigma = 0.01
Noise-conditioning scalar for the EDMUNet instantiation during demosaicing.

assumptions (4)

domain assumption Instance segmentation supervision on COCO produces region- and boundary-aware decoder features that are useful for demosaicing.
Central transfer premise of the paper; introduced in Sections 1 and 3.2.
domain assumption DINOv2 features extracted from rendered sRGB images provide a valid shared structural reference for both segmentation and demosaicing decoder features.
Used in Section 3.4 as the anchoring target for both tasks.
domain assumption The fixed inverse PTP transform (fixed camera profile, gamma 2.2, fixed 3x3 color correction matrix) is a valid model for converting display sRGB images to camera-linear RGB for training and evaluation.
Used for synthetic dataset construction in Appendix A and for all training data.
standard math The CFA measurement model y = M_p(x) + n with Gaussian sensor noise is the standard forward model for demosaicing.
Introduced in Eq. (1) and used throughout.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SegDem: Segmentation helps Demosaicing." pith.science (2026). https://pith.science/paper/73EPEB7Y

@misc{pith2026260807916,
  author       = {Pith},
  title        = {Pith review of: SegDem: Segmentation helps Demosaicing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/73EPEB7Y}},
  note         = {Machine review of arXiv:2608.07916}
}

read the original abstract

Image demosaicing reconstructs a full-color image from incomplete color measurements produced by a sensor covered with a color filter array (CFA). Most existing methods formulate demosaicing as pixel-level reconstruction and mainly rely on local textures, cross-channel correlations, and low-level image statistics. Our core insight is that reconstruction and visual understanding can be viewed as complementary views of shared scene structure: both are grounded in the same underlying physical world, and therefore the structural and physical information inferred from an image should remain consistent across the two tasks. We instantiate this idea with instance segmentation and propose \emph{SegDem}, a cross-task decoder representation transfer framework for demosaicing. SegDem first learns region- and boundary-aware representations through instance-aware structural pretraining and then transfers the decoder to RAW-conditioned reconstruction. Segmentation- and demosaicing-conditioned features are further anchored to a shared frozen DINOv2 representation space to preserve structural organization across tasks. We instantiate SegDem with convolutional, Transformer-based, and state-space backbones for unified Single- and Quad-Bayer demosaicing. Extensive experiments on synthetic, external, and challenging datasets demonstrate consistent improvements across different architectures and CFA layouts.

Figures

Figures reproduced from arXiv: 2608.07916 by the authors.

**Figure 1.** Motivation of SegDem. Conventional regression is prone to false colors and moire arti- ´ facts in ambiguous regions, while generative models may hallucinate details unsupported by RAW measurements. SegDem instead transfers structural decoder representations from segmentation to demosaicing, improving boundary-aware reconstruction. ABSTRACT Image demosaicing reconstructs a full-color image from incomplete color meas… view at source ↗

**Figure 2.** Overview of SegDem. Stage 1 learns region- and boundary-aware representations in a shared RAW-conditioned decoder. Stage 2 transfers the pretrained decoder parameters to demosaicing, while task-conditioned auxiliary features and training-only DINOv2 alignment provide structural supervision without altering the sensor-conditioned output pathway. channel correlation priors (Malvar et al., 2004). Learning-based method… view at source ↗

**Figure 3.** Motivation for DINO-bridged structural anchoring. Segmentation- and demosaicingconditioned decoder tokens are independently aligned to the same frozen DINOv2 representation space, providing a common structural reference without directly matching task-specific features. Scratch Seg. Pretrained SegDm Full Scratch Seg. Pretrained SegDm Full Scratch Seg. Pretrained SegDm Full Scratch Seg. Pretrained SegDm Full [PITH_F… view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Decoder representation analysis. Scratch training yields local boundary responses but limited separation between regions. Structural pretraining produces more region-discriminative decoder features, and the full model largely preserves this organization after demosaic…

**Figure 5.** Figure 5: Synthetic visual comparison. Each row uses the same mosaicked observation and compares final demosaicing predictions from representative baselines and SegDem. PSNR/SSIM results, while EDMUNet gives the lowest LPIPS. This difference is consistent with different backbon…

**Figure 6.** Figure 6: MIT-Hard visual comparison. Each example shows the mosaicked input followed by the final sRGB prediction of each model on the same high-frequency observation [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Visual ablation of structural pretraining. The error maps highlight regions where decoder transfer reduces structured color artifacts, especially around boundaries and repetitive textures [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: shows the detailed condition and representation-target construction used by SegDem. The two branches serve different roles. The sensor condition branch builds a measurement-aware RAW condition for linear RGB reconstruction, while the representation target branch builds…

**Figure 9.** Figure 9: illustrates the instruction adapter used to make the shared decoder representation taskaware. The adapter is intentionally lightweight. It does not concatenate language tokens with the RAW input and does not introduce an autoregressive text-to-image generation path. I…

**Figure 10.** Figure 10: Supplementary segmentation visual comparison. From left to right, we show the rendered input, ground-truth instance masks, segmentation-only prediction, and demosaicing-initialized prediction [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 54 canonical work pages

[1]

Scaling Learning Algorithms Towards

Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards
[2]

and Osindero, Simon and Teh, Yee Whye , journal =

Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =
[3]

2016 , publisher=

Deep learning , author=. 2016 , publisher=

2016
[4]

, title =

Bayer, Bryce E. , title =. 1976 , note =

work page 1976
[5]

and He, Li-wei and Cutler, Ross , title =

Malvar, Henrique S. and He, Li-wei and Cutler, Ross , title =. IEEE International Conference on Acoustics, Speech, and Signal Processing , year =

work page
[6]

Deep joint demosaicking and denoising , journal =

Gharbi, Micha. Deep joint demosaicking and denoising , journal =

work page
[7]

2018 , eprint =

Kokkinos, Filippos and Lefkimmiatis, Stamatios , title =. 2018 , eprint =

work page 2018
[8]

IEEE/CVF International Conference on Computer Vision Workshops , year =

Liang, Jingyun and Cao, Jiezhang and Sun, Guolei and Zhang, Kai and Van Gool, Luc and Timofte, Radu , title =. IEEE/CVF International Conference on Computer Vision Workshops , year =

work page

Show all 61 references

[9]

IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

Zamir, Syed Waqas and Arora, Aditya and Khan, Salman and Hayat, Munawar and Khan, Fahad Shahbaz and Yang, Ming-Hsuan , title =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =
[10]

2024 , eprint =

Guo, Hang and Li, Jinmin and Dai, Tao and Ouyang, Zhihao and Ren, Xudong and Xia, Shu-Tao , title =. 2024 , eprint =

2024
[11]

2023 , eprint =

Lee, Haechang and Park, Dongwon and Jeong, Wongi and Kim, Kijeong and Je, Hyunwoo and Ryu, Dongil and Chun, Se Young , title =. 2023 , eprint =

2023
[12]

2025 , eprint =

Zhou, Shiyang and Zeng, Haijin and Lu, Yunfan and Chen, Yongyong and Liu, Jie and Su, Jingyong , title =. 2025 , eprint =

2025
[13]

2026 , eprint =

Kumar, Saurabh and Yenneti, Nutan Sairam , title =. 2026 , eprint =

2026
[14]

Wang, Jianyi and Yue, Zongsheng and Zhou, Shangchen and Chan, Kelvin C. K. and Loy, Chen Change , title =. IEEE/CVF International Conference on Computer Vision , year =
[15]

2023 , eprint =

Lin, Xinqi and He, Jingwen and Chen, Ziyan and Lyu, Zhaoyang and Dai, Bo and Yu, Fanghua and Ouyang, Wanli and Qiao, Yu and Dong, Chao , title =. 2023 , eprint =

2023
[16]

2023 , eprint =

Yang, Tao and Wu, Rongyuan and Ren, Peiran and Xie, Xuansong and Zhang, Lei , title =. 2023 , eprint =

2023
[17]

IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

Wu, Rongyuan and Sun, Lingchen and Ma, Zhiyuan and Zhang, Lei , title =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =
[18]

and Gu, J

Yu, F. and Gu, J. and Li, Z. and Hu, J. and Kong, X. and Wang, X. and He, J. and Qiao, Y. and Dong, C. , title =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =
[19]

2024 , eprint =

Wu, Rongyuan and Sun, Lingchen and Ma, Zhiyuan and Zhang, Lei , title =. 2024 , eprint =

2024
[20]

IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

Blau, Yochai and Michaeli, Tomer , title =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =
[21]

2026 , eprint =

Yin, Xiang and Hu, Jinfan and You, Zhiyuan and Yan, Kainan and Tang, Yu and Dong, Chao and Gu, Jinjin , title =. 2026 , eprint =

2026
[22]

He, Kaiming and Gkioxari, Georgia and Doll. Mask. IEEE International Conference on Computer Vision , year =
[23]

2023 , eprint =

Oquab, Maxime and Darcet, Timoth. 2023 , eprint =

2023
[24]

and Zhao, Zheng and Sj

Luo, Ziwei and Gustafsson, Fredrik K. and Zhao, Zheng and Sj. Controlling Vision-Language Models for Multi-Task Image Restoration , year =. 2310.01018 , archivePrefix =

arXiv
[25]

2025 , eprint =

Yang, Cuixin and Dong, Rongkang and Lam, Kin-Man , title =. 2025 , eprint =

2025
[26]

2026 , eprint =

Guan, Kai and Wu, Rongyuan and Li, Shuai and Zhu, Wentao and Zeng, Wenjun and Zhang, Lei , title =. 2026 , eprint =

2026
[27]

2026 , eprint =

Chen, Ping and Huang, Zicheng and Wang, Xiangming and Liu, Yungeng and Liang, Bingyu and Zeng, Haijin and Chen, Yongyong , title =. 2026 , eprint =

2026
[28]

CoRR , volume =

Yufeng Yang and Xianfang Zeng and Zhangqi Jiang and Fukun Yin and Jianzhuang Liu and Wei Cheng and Jinghong Lan and Shiyu Liu and Yuqi Peng and Gang Yu and Shifeng Chen , title =. CoRR , volume =
[29]

CoRR , volume =

Weixiong Sun and Xiang Yin and Chao Dong , title =. CoRR , volume =
[30]

Brown , title =

SaiKiran Kumar Tedla and Abhijith Punnappurath and Luxi Zhao and Michael S. Brown , title =
[31]

Brown , title =

SaiKiran Kumar Tedla and Junyong Lee and Beixuan Yang and Mahmoud Afifi and Michael S. Brown , title =
[32]

Shiyang Zhou and Haijin Zeng and Yunfan Lu and Tong Shao and Ke Tang and Yongyong Chen and Jie Liu and Jingyong Su , title =
[33]

The Thirteenth International Conference on Learning Representations,

Sihyun Yu and Sangkyung Kwak and Huiwon Jang and Jongheon Jeong and Jonathan Huang and Jinwoo Shin and Saining Xie , title =. The Thirteenth International Conference on Learning Representations,
[34]

CoRR , volume =

Jaskirat Singh and Xingjian Leng and Zongze Wu and Liang Zheng and Richard Zhang and Eli Shechtman and Saining Xie , title =. CoRR , volume =
[35]

CoRR , volume =

Nilaksh and Saurav Jha and Artem Zholus and Sarath Chandar , title =. CoRR , volume =
[36]

CoRR , volume =

Ping Chen and Zicheng Huang and Xiangming Wang and Yungeng Liu and Bingyu Liang and Haijin Zeng and Yongyong Chen , title =. CoRR , volume =. 2026 , eprinttype =

2026
[37]

Advances in Neural Information Processing Systems , volume =

Tero Karras and Miika Aittala and Timo Aila and Samuli Laine , title =. Advances in Neural Information Processing Systems , volume =
[38]

Feature Pyramid Networks for Object Detection , booktitle =

Tsung-Yi Lin and Piotr Doll. Feature Pyramid Networks for Object Detection , booktitle =
[39]

arXiv preprint arXiv:2502.13923 , year =

Shuai Bai and Keqin Chen and Xuejing Liu and Jialin Wang and Wenbin Ge and Sibo Song and Kai Dang and Peng Wang and Shijie Wang and Jun Tang and Humen Zhong and Yuanzhi Zhu and Mingkun Yang and Zhaohai Li and Jianqiang Wan and Pengfei Wang and Wei Ding and Zheren Fu and Yiheng...

arXiv
[40]

Yawei Li and Kai Zhang and Jingyun Liang and Jiezhang Cao and Ce Liu and Rui Gong and Yulun Zhang and Hao Tang and Yun Liu and Denis Demandolx and Rakesh Ranjan and Radu Timofte and Luc Van Gool , title =
[41]

European Conference on Computer Vision (ECCV) , pages=

Microsoft COCO: Common Objects in Context , author=. European Conference on Computer Vision (ECCV) , pages=. 2014 , organization=

2014
[42]

Mathematics , VOLUME =

Zhang, Donghui and Li, Feiyu and Yang, Jun and Yang, Le , TITLE =. Mathematics , VOLUME =. 2026 , NUMBER =

2026
[43]

Haijin Zeng and Kai Feng and Jiezhang Cao and Shaoguang Huang and Yongqiang Zhao and Hiep Quang Luong and Jan Aelterman and Wilfried Philips , title =. Int. J. Comput. Vis. , volume =
[44]

Haijin Zeng and Xiangming Wang and Yongyong Chen and Jingyong Su and Jie Liu , title =
[45]

High-Resolution Image Synthesis with Latent Diffusion Models , booktitle =

Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and Bj. High-Resolution Image Synthesis with Latent Diffusion Models , booktitle =
[46]

CoRR , volume =

Black Forest Labs and Stephen Batifol and Andreas Blattmann and Frederic Boesel and Saksham Consul and Cyril Diagne and Tim Dockhorn and Jack English and Zion English and Patrick Esser and Sumith Kulal and Kyle Lacey and Yam Levi and Cheng Li and Dominik Lorenz and Jonas M. Co...
[47]

Learning Transferable Visual Models From Natural Language Supervision , booktitle =

Alec Radford and Jong Wook Kim and Chris Hallacy and Aditya Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and Pamela Mishkin and Jack Clark and Gretchen Krueger and Ilya Sutskever , editor =. Learning Transferable Visual Models From Natural La...
[48]

Courville , editor =

Ethan Perez and Florian Strub and Harm de Vries and Vincent Dumoulin and Aaron C. Courville , editor =. FiLM: Visual Reasoning with a General Conditioning Layer , booktitle =
[49]

Proceedings of the AAAI Conference on Artificial Intelligence , pages =

Zhang, Tao and Fu, Ying and Li, Cheng , title =. Proceedings of the AAAI Conference on Artificial Intelligence , pages =
[50]

Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

You, Weiyi and Zhang, Mingyang and Zhang, Leheng and Zhou, Xingyu and Shi, Kexuan and Gu, Shuhang , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

2025
[51]

Xiangming Wang and Haijin Zeng and Benteng Sun and Jiezhang Cao and Kai Zhang and Qiangqiang Shen and Yongyong Chen , title =
[52]

CoRR , volume =

Xiangming Wang and Benteng Sun and Yungeng Liu and Haijin Zeng and Yongyong Chen and Jingyong Su and Jie Liu , title =. CoRR , volume =
[53]

CoRR , volume =

Om Khangaonkar and Hamed Pirsiavash , title =. CoRR , volume =. 2025 , eprinttype =

2025
[54]

Rabbat and Yann LeCun and Nicolas Ballas , title =

Mahmoud Assran and Quentin Duval and Ishan Misra and Piotr Bojanowski and Pascal Vincent and Michael G. Rabbat and Yann LeCun and Nicolas Ballas , title =
[55]

Connecting Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learning , booktitle =

Shentong Mo and Peter Tong , editor =. Connecting Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learning , booktitle =
[56]

arXiv preprint arXiv:2607.26600 , year=

JEPADepth: Predictive Representation Learning for Self-Supervised Monocular Depth Estimation , author=. arXiv preprint arXiv:2607.26600 , year=

arXiv
[57]

CoRR , volume =

Yan Chen and Yi Wen and Wei Li and Junchao Liu and Yong Guo and Jie Hu and Xinghao Chen , title =. CoRR , volume =. 2025 , eprinttype =

2025
[58]

The Thirteenth International Conference on Learning Representations,

Haijin Zeng and Benteng Sun and Yongyong Chen and Jingyong Su and Yong Xu , title =. The Thirteenth International Conference on Learning Representations,
[59]

MambaSCI: Efficient Mamba-UNet for Quad-Bayer Patterned Video Snapshot Compressive Imaging , booktitle =

Zhenghao Pan and Haijin Zeng and Jiezhang Cao and Yongyong Chen and Kai Zhang and Yong Xu , editor =. MambaSCI: Efficient Mamba-UNet for Quad-Bayer Patterned Video Snapshot Compressive Imaging , booktitle =
[60]

Wavelength-Embedding-Guided Filter-Array Transformer for Spectral Demosaicing , booktitle =

Haijin Zeng and Hiep Luong and Wilfried Philips , editor =. Wavelength-Embedding-Guided Filter-Array Transformer for Spectral Demosaicing , booktitle =
[61]

Deep Joint Demosaicking and Denoising , journal =

Gharbi, Micha\". Deep Joint Demosaicking and Denoising , journal =. 2016 , issn =

2016

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Scaling Learning Algorithms Towards

Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

[2] [2]

and Osindero, Simon and Teh, Yee Whye , journal =

Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

[3] [3]

2016 , publisher=

Deep learning , author=. 2016 , publisher=

2016

[4] [4]

, title =

Bayer, Bryce E. , title =. 1976 , note =

work page 1976

[5] [5]

and He, Li-wei and Cutler, Ross , title =

Malvar, Henrique S. and He, Li-wei and Cutler, Ross , title =. IEEE International Conference on Acoustics, Speech, and Signal Processing , year =

work page

[6] [6]

Deep joint demosaicking and denoising , journal =

Gharbi, Micha. Deep joint demosaicking and denoising , journal =

work page

[7] [7]

2018 , eprint =

Kokkinos, Filippos and Lefkimmiatis, Stamatios , title =. 2018 , eprint =

work page 2018

[8] [8]

IEEE/CVF International Conference on Computer Vision Workshops , year =

Liang, Jingyun and Cao, Jiezhang and Sun, Guolei and Zhang, Kai and Van Gool, Luc and Timofte, Radu , title =. IEEE/CVF International Conference on Computer Vision Workshops , year =

work page