Pith. sign in

REVIEW 5 major objections 5 minor 63 references

Implicit text prompts let a frozen image-editing VLM perform a task B from a single demonstration pair of task A, without naming either task.

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 · deepseek-v4-flash

2026-08-03 21:11 UTC pith:37NVSLZZ

load-bearing objection A genuinely new cross-task VICL framing and dataset, but the evaluation is compromised by test-time PSNR selection against ground truth. the 5 major comments →

arxiv 2511.16107 v4 pith:37NVSLZZ submitted 2025-11-20 cs.CV cs.AI

T2T-VICL: Cross-Task Visual In-Context Learning via Implicit Text-Driven VLMs

Shao-Jun Xia , Huixin Zhang , Zhengzhong Tu This is my paper
classification cs.CV cs.AI
keywords visual in-context learningcross-task transfervision-language modelsimplicit text promptingknowledge distillationimage editinglow-level visionVIEScore
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.

This paper tries to establish that visual in-context learning can work across different tasks: given a single input-output example from task A and a query image from task B, a frozen image-editing vision-language model can produce task-B output, provided the difference between the two tasks is captured as an implicit text prompt. To get that prompt, a large teacher VLM generates structured descriptions of the visual changes between task pairs, and a lightweight student VLM is distilled to produce content-dependent prompts on demand. The authors build a dataset of implicit cross-task descriptions, couple this with a score-based inference scheme, and report consistent gains in task-aware alignment over fixed prompting across 21 task pairs. If correct, this would mean task knowledge can be transferred between vision tasks in a single pass, without retraining, by translating visual differences into language.

Core claim

The central claim is that cross-task VICL is possible: a vision-language model can execute a target task B from a query image and a single input-output demonstration of a different task A, as long as the relationship between the two tasks is expressed in an implicit, content-dependent text prompt. The prompt is produced by a student VLM that was distilled from a larger teacher VLM's structured descriptions of visual changes and task differences, and it is consumed by a frozen image-editing VLM to generate the output. The paper shows this on 12 low-level vision tasks and 21 cross-task pairs, with top-tier VIEScores in twelve cases and second-tier VIEScores in the remaining nine, and argues th

What carries the argument

The central mechanism is the implicit text prompt (the 'implicit relations' description) generated by a distilled student vision-language model from the task-A demonstration pair and task-B query, then fed to a frozen image-editing VLM to guide generation. The student is trained on teacher-generated, diversity-filtered textual comparisons that describe visual changes and task differences without naming the tasks. This prompt-transfer loop is supported by a score-based inference step that runs the frozen model multiple times and selects the candidate with the highest PSNR against the ground-truth target, with VIEScore used as a task-aware evaluation metric.

Load-bearing premise

The head-to-head comparison assumes that the 10-run PSNR-based selection in Sec 3.3.4 is a fair and valid evaluation protocol for both T2T-VICL and the fixed-prompt baseline; if the baseline is not given the same selection, the claimed consistent improvements may be an artifact of test-time fitting to the ground truth rather than a property of the implicit prompts.

What would settle it

Run both T2T-VICL and the fixed-prompt baseline on a held-out set of at least 100 queries per task pair, giving both methods exactly 10 generated candidates and selecting the best by PSNR; if the VIEScore and PSNR gains over fixed prompting fall to near zero, the central claim of cross-task prompt-driven transfer would be falsified.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If cross-task VICL holds, users could apply a frozen image-editing model to a new transformation using only one example from a related task, with no fine-tuning or task-specific code.
  • The teacher-student distillation suggests that task-relationship knowledge can be compressed into a lightweight model that generates prompts on the fly, making deployment cheaper.
  • The consistent VIEScore gains imply that task-aware semantic alignment improves across task boundaries, even when pixel-level fidelity is unchanged or slightly lower.
  • The method extends to both intra-category and inter-category task pairs, indicating that the implicit-text prompt is not limited to visually similar tasks.
  • The work provides a first dataset of implicit cross-task descriptions, which could support future research on cross-task generalization in multimodal models.

Where Pith is reading between the lines

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

  • A testable extension would be to use the same implicit-prompt pipeline for non-image modalities (e.g., audio or video) where transformations can also be described in language, though the paper does not explore this.
  • The reported gains depend on selecting the best of 10 candidates by ground-truth PSNR; in a real deployment without ground truth, a no-reference selector (e.g., VIEScore alone) would be needed, and the advantage over fixed prompting may shrink.
  • If the fixed-prompt baseline were also given 10 generated candidates with the same PSNR-oracle selection, the gap between the methods could narrow or disappear, so the head-to-head comparison is a weaker form of evidence than the paper's tone suggests.
  • The implicit prompt may be doing more than describing task differences: it could also serve as a compositional instruction that lets users combine known tasks to specify novel transformations, a direction the paper leaves unstated.

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

5 major / 5 minor

Summary. The paper introduces T2T-VICL, a pipeline for cross-task visual in-context learning (VICL) in which a large teacher VLM (Qwen2.5-VL-32B) generates implicit textual descriptions of the differences between pairs of low-level vision tasks; a student VLM (Qwen2.5-VL-4B) is fine-tuned to produce such descriptions from a demonstration pair and a query; and the resulting prompt is fed to a frozen Gemini-2.5-Flash model that edits the query image. Candidate outputs from ten inference runs are selected by picking the one with the highest PSNR against the ground-truth target, and PSNR, SSIM, and VIEScore are reported for 12 top-tier and 9 second-tier cross-task pairs. The paper claims consistent task-aware alignment gains over a fixed-prompt baseline, with mixed fidelity results.

Significance. If the evaluation were valid, this would be a useful contribution: it formulates a genuinely cross-task VICL setting, constructs a large implicit-description dataset, and demonstrates a teacher-student prompt-generation mechanism with a frozen image-editing VLM. The breadth of tasks (12 low-level tasks) and the public code release are strengths. However, the central comparison is compromised by the test-time PSNR oracle in §3.3.4, so the reported improvements cannot currently be attributed to the proposed method. The dataset is also self-generated by the teacher with no human validation, and the metric used for the main claim (VIEScore) is itself VLM-based. The ideas are promising, but the evidence needs substantial rework.

major comments (5)
  1. [§3.3.4, Tables 2–3] The evaluation protocol selects, for each query, the inference candidate with the highest PSNR against the ground-truth target: 'we performed inference 10 times and selected the result with the highest PSNR as the final output while computing the corresponding VIEScore and SSIM.' This is a test-time oracle: PSNR requires the ground-truth image, which is unavailable in any real deployment. All reported PSNR, SSIM, and VIEScore numbers are best-of-10 values. The paper never states whether the fixed-prompt baseline receives the same selection; if it does not, Tables 2 and 3 compare an oracle-selected method against a single-sample baseline. Even if the baseline is also oracle-selected, the reported numbers are upper envelopes, not expected performance, and no mean/standard deviation over the 10 runs is provided. The consistent-improvement claim therefore cannot be attributed to the learned
  2. [§3.3.3–3.3.4] There is a mismatch between the proposed inference mechanism and the evaluated protocol. §3.3.3 introduces 'perceptual score-based screening based on VIEScore' as the ranking mechanism, and Fig. 2 shows candidate ranking with IQA metrics. However, §3.3.4 states that selection is performed by PSNR, not VIEScore. The paper does not clarify whether VIEScore participates in selection at all, and the score-based reasoning described in Eqs. (5)–(6) is not what produces the reported outputs. The reader is left unsure whether the method's contribution is the learned prompts or the PSNR oracle. Please specify the selection procedure precisely and evaluate the method under its own design (e.g., VIEScore-based selection, or a fixed candidate without oracle).
  3. [§3.2, §3.3.1, §4.2] The implicit-description dataset is generated by the teacher VLM (Qwen-32B) and the student is trained to imitate that teacher. No human validation or independent evaluation of the 2,000 descriptions per task pair is reported. The claim that these descriptions 'implicitly capture cross-task relationships' rests entirely on the teacher's own output. Moreover, the main 'task-aware alignment' metric, VIEScore, is itself driven by a VLM; using a VLM metric to evaluate prompts produced by a VLM of the same family creates a risk of shared bias. This is not a definitional circularity because final images are compared to ground truth, but it does weaken the evidence that the prompts are semantically meaningful and that the VIEScore gains reflect true task alignment rather than a preference for teacher-style language.
  4. [Tables 2–3, §4.5] Even under the favorable oracle protocol, the data do not consistently support the headline claim. In Table 3, PSNR declines in 7 of 9 rows and SSIM declines in 6 of 9 rows; in Table 2, SSIM declines in two rows and several PSNR gains are small. The paper acknowledges 'slight declines' in §4.5 but provides no error bars, significance tests, or per-run distributions. With 10 repeated inferences per query, the variance across candidates is readily available and should be reported. Without this, it is impossible to judge whether the observed VIEScore improvements and PSNR differences are meaningful or artifacts of the oracle selection.
  5. [§3.3.1–3.3.2] The paper does not ablate the student component. The central design is a teacher-student distillation, but there is no comparison of the proposed T2T-VICL against (a) directly using the teacher VLM's implicit descriptions at inference time, or (b) using a fixed task-naming prompt (e.g., 'denoise this image') instead of the implicit description. Without such ablations, the contribution of the lightweight student model and of the 'implicit' formulation is not established. At minimum, a teacher-prompt baseline would clarify whether the student adds value or merely approximates the teacher.
minor comments (5)
  1. [§3.3.2] The text says 'the Small→Large step uses the 7B model' but the student is Qwen2.5-VL-4B; this is inconsistent. Also, the notation P'_t, P''_t, and P_s in Eqs. (2)–(4) is confusing; please use distinct symbols for the student-generated prompt and the final prompt passed to the large VLM.
  2. [Fig. 2] Panel (C) shows '×10 rounds' and 'IQA Metrics PSNR SSIM' but does not show VIEScore in the selection loop, although §3.3.3 describes VIEScore-based reasoning. Please align the figure with the actual selection mechanism.
  3. [§4.1] No test-set sizes are reported for any of the 12 tasks or for the 21 cross-task pairs. The number of query images evaluated is essential for interpreting the averages in Tables 2 and 3 and for any future reproducibility check.
  4. [§4.2] The sentence 'at least two of the three evaluation metrics consistently outperform the fixed prompt baseline' is vague. Please identify explicitly which metrics improve and which do not, perhaps with a per-task breakdown in the tables.
  5. [Abstract and §4.6] The terms 'top-tier' and 'second-tier' are not defined. It appears they refer to the two tables, but the wording could suggest a ranking of methods. Please define these terms operationally.

Circularity Check

1 steps flagged

PSNR-oracle selection in §3.3.4 makes the reported outputs fitted to ground truth; the headline comparison reduces to test-time selection rather than the learned prompts.

specific steps
  1. fitted input called prediction [Section 3.3.4 'Evaluation metrics'; results in Tables 2 and 3 and Figure 3]
    "For each query image, we performed inference 10 times and selected the result with the highest PSNR as the final output while computing the corresponding VIEScore and SSIM."

    The reported 'prediction' is defined as the sample with maximum PSNR against the ground-truth target of task B—a test-time oracle unavailable in deployment. All PSNR/SSIM/VIEScore numbers in Tables 2–3 are therefore properties of the argmax-over-GT selector, not of the implicit prompt or frozen VLM. The central claim ('consistently improves task-aware alignment over fixed prompting') is supported only by this oracle-selected comparison, so the improvement reduces to test-time fitting. The paper does not state whether the baseline receives the same selection and reports no non-oracle means, so the head-to-head evidence has no independent content beyond the selector.

full rationale

The only step that reduces by construction is the evaluation oracle in §3.3.4. The teacher-student distillation loop (§3.2–3.3.1) is not itself circular: the student is trained to imitate the teacher's text, but the final image outputs are scored against external ground-truth images and an external metric (VIEScore), so the central claim is not definitionally forced by the training data. There is no load-bearing self-citation chain; SPIRE [34] is cited only as related work and no uniqueness theorem is imported. The claimed contribution of implicit prompt transfer therefore has independent content in principle. However, the reported head-to-head superiority over fixed prompting is established exclusively through the §3.3.4 protocol of running inference 10 times and keeping the candidate with the highest PSNR to the ground truth. Because that selector requires the task-B label, the 'prediction' is fitted to the test target, making the improvement a property of the oracle rather than of the learned prompts; additionally the VIEScore-based reasoning described in §3.3.3 is never the selection mechanism actually evaluated. This warrants a partial circularity score of 6.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper is an empirical ML pipeline, so the ledger mostly records hand-chosen hyperparameters and untested domain assumptions about VLM prompt-transfer rather than fitted constants.

free parameters (3)
  • implicit descriptions per task pair = 2000
    Hand-chosen dataset size; no sensitivity analysis; affects student training diversity.
  • inference candidates per query (k) = 10
    Ten stochastic generations per query; final selected by highest PSNR; no analysis of k or selection metric.
  • student VLM size = Qwen2.5-VL-4B-Instruct
    Choice of 4B student vs other sizes; no ablation reported.
axioms (4)
  • domain assumption A large VLM (Qwen2.5-VL-32B) can produce reliable implicit comparisons of task pairs from images and labels.
    Sec 3.2 uses teacher outputs as ground truth for dataset and student supervision; no human validation reported, only diversity filtering.
  • domain assumption The distilled 4B student can generate content-dependent prompts that, when fed to Gemini-2.5-Flash, transfer the task-B transformation.
    Sec 3.3.2 assumes prompt transfer from small to large VLM; no direct evaluation of prompt quality.
  • domain assumption VIEScore is a valid measure of task-aware alignment for cross-task outputs.
    Sec 3.3.3 relies on VIEScore as explainable evaluator; improvement in VIEScore is the main evidence for task-aware alignment.
  • domain assumption The 12 public datasets provide comparable test distributions for cross-task pairing.
    Sec 4.1 pairs different dataset domains; no control for domain shift.

pith-pipeline@v1.3.0-alltime-deepseek · 14344 in / 10307 out tokens · 92836 ms · 2026-08-03T21:11:41.293922+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of T2T-VICL: Cross-Task Visual In-Context Learning via Implicit Text-Driven VLMs." pith.science (2026). https://pith.science/paper/37NVSLZZ

@misc{pith2026251116107,
  author       = {Pith},
  title        = {Pith review of: T2T-VICL: Cross-Task Visual In-Context Learning via Implicit Text-Driven VLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/37NVSLZZ}},
  note         = {Machine review of arXiv:2511.16107}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Visual in-context learning (VICL) solves visual tasks by conditioning on a few input-output demonstrations without any model training. Recent advances in large vision-language models (VLMs) have shown promising VICL capability when the demonstration pair and the query belong to the same vision task, but real use cases often provide mismatched examples, making it unclear whether a VLM should imitate the demonstrated transformation or infer a new one from the query. This raises a fundamental question: Can VLMs perform cross-task VICL where demonstration and query differ? In the paper, we study this cross-task VICL setting and propose T2T-VICL, a collaborative prompt-transfer framework, which converts mismatched visual demonstrations into implicit textual guidance without explicitly naming the tasks. To do so, a large teacher VLM first generates structured descriptions of visual changes and task differences between task pairs, from which we construct a dataset of diverse implicit cross-task relations. We then distill this capability into a lightweight student VLM that produces content-dependent prompts from a task-A demonstration pair and a task-B query. The generated prompt is used to guide a frozen image-editing VLM, and a score-based inference strategy is introduced to rank multiple candidates. Experiments on 12 low-level vision tasks and over 20 evaluated cross-task pairs show that T2T-VICL consistently improves task-aware alignment over fixed prompting and often also improves image fidelity, revealing both the potential and limits of cross-task VICL. Our code is available on GitHub.

Figures

Figures reproduced from arXiv: 2511.16107 by Huixin Zhang, Shao-Jun Xia, Zhengzhong Tu.

Figure 1
Figure 1. Figure 1: Illustration of Cross-Task Visual In-Context Learning [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Proposed Pipeline and its Workflow [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative Examples of Cross-Task In-Context Learning in Twelve Pairs [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative Examples of Cross-Task In-Context Learning from the Nine Pairs [PITH_FULL_IMAGE:figures/full_fig_p008_4.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

63 extracted references · 6 linked inside Pith

  1. [1]

    Abdelrahman Abdelhamed, Stephen Lin, and Michael S. Brown. A high-quality denoising dataset for smartphone cameras. InIEEE Conference on Computer Vision and Pat- tern Recognition, CVPR 2018, pages 1692–1700. Computer Vision Foundation / IEEE Computer Society, 2018. 6

    2018

  2. [2]

    Abdelrahman Abdelhamed, Mahmoud Afifi, Radu Timo- fte, Michael S. Brown, Yue Cao, Zhilu Zhang, Wangmeng Zuo, Xiaoling Zhang, Jiye Liu, Wendong Chen, Changyuan Wen, Meng Liu, Shuailin Lv, Yunchao Zhang, Zhihong Pan, Baopu Li, Teng Xi, Yanwen Fan, Xiyu Yu, Gang Zhang, Jingtuo Liu, Junyu Han, Errui Ding, Songhyun Yu, Bumjun Park, Jechang Jeong, Shuai Liu, Zi...

    2020

  3. [3]

    Fowlkes, Ste- fano Soatto, and Pietro Perona

    Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C. Fowlkes, Ste- fano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. InIEEE/CVF International Conference on Computer Vision, ICCV 2019, pages 6429–6438. IEEE,

    2019

  4. [4]

    NTIRE 2017 chal- lenge on single image super-resolution: Dataset and study

    Eirikur Agustsson and Radu Timofte. NTIRE 2017 chal- lenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recogni- tion Workshops, CVPR Workshops 2017, pages 1122–1131. IEEE Computer Society, 2017. 6

    2017

  5. [5]

    Ancuti, and Christophe De Vleeschouwer

    Cosmin Ancuti, Codruta O. Ancuti, and Christophe De Vleeschouwer. D-HAZY: A dataset to evaluate quantita- tively dehazing algorithms. InIEEE International Confer- ence on Image Processing, ICIP 2016, pages 2226–2230. IEEE, 2016. 6

    2016

  6. [6]

    Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas J. Guibas. An information-theoretic approach to transferability in task transfer learning. InIEEE International Conference on Im- age Processing, ICIP 2019, pages 2309–2313. IEEE, 2019. 1

    2019

  7. [7]

    Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Glober- son, and Alexei A. Efros. Visual prompting via image in- painting. InAdvances in Neural Information Processing Sys- tems 35: Annual Conference on Neural Information Process- ing Systems 2022, NeurIPS 2022, 2022. 1, 2

    2022

  8. [8]

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros. In- structpix2pix: Learning to follow image editing instructions. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, pages 18392–18402. IEEE, 2023. 1, 3

    2023

  9. [9]

    Fleet, and Geoffrey E

    Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J. Fleet, and Geoffrey E. Hinton. A unified sequence interface for vision tasks. InAdvances in Neural Information Process- ing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, 2022. 2

    2022

  10. [10]

    Lovish Chum, Anbumani Subramanian, Vineeth N Balasub- ramanian, and C.V . Jawahar. Beyond supervised learning: A computer vision perspective.Journal of the Indian Institute of Science, 99(2):177–199, 2019. 1

    2019

  11. [11]

    Conde, Gregor Geigle, and Radu Timofte

    Marcos V . Conde, Gregor Geigle, and Radu Timofte. Instruc- tir: High-quality image restoration following human instruc- tions. InProceedings of the 18th European Conference on Computer Vision, ECCV 2024, pages 1–21. Springer, 2024. 1, 3

    2024

  12. [12]

    Image harmonization dataset iharmony4: Hcoco, hadobe5k, hflickr, and hday2night

    Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Image harmonization dataset iharmony4: Hcoco, hadobe5k, hflickr, and hday2night. CoRR, abs/1908.10526, 2019. 6

    Pith/arXiv arXiv 1908

  13. [13]

    A survey on in-context learning

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, Xu Sun, Lei Li, and Zhifang Sui. A survey on in-context learning. InProceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Processing, EMNLP 2024, pages 1107–1128. Association for Computa- tional Linguistics, 2024. 1

    2024

  14. [14]

    Representation similar- ity analysis for efficient task taxonomy & transfer learning

    Kshitij Dwivedi and Gemma Roig. Representation similar- ity analysis for efficient task taxonomy & transfer learning. InIEEE Conference on Computer Vision and Pattern Recog- nition, CVPR 2019, pages 12387–12396. Computer Vision Foundation / IEEE, 2019. 3

    2019

  15. [15]

    Yulu Gan, Sungwoo Park, Alexander Schubert, Anthony Philippakis, and Ahmed M. Alaa. Instructcv: Instruction- tuned text-to-image diffusion models as vision generalists. In Proceedings of the 12th International Conference on Learn- ing Representations, ICLR 2024. OpenReview.net, 2024. 2

    2024

  16. [16]

    Instructdiffusion: A generalist modeling interface for vision tasks

    Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, Dong Chen, and Baining Guo. Instructdiffusion: A generalist modeling interface for vision tasks. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, pages 12709–12720. IEEE, 2024. 2

    2024

  17. [17]

    In-context learning in large language mod- els: A comprehensive survey

    Clyde Highmore. In-context learning in large language mod- els: A comprehensive survey. 2024. 1

    2024

  18. [18]

    Depth-attentional features for single-image rain removal

    Xiaowei Hu, Chi-Wing Fu, Lei Zhu, and Pheng-Ann Heng. Depth-attentional features for single-image rain removal. In IEEE Conference on Computer Vision and Pattern Recogni- tion, CVPR 2019, pages 8022–8031. Computer Vision Foun- dation / IEEE, 2019. 6

    2019

  19. [19]

    Belongie, Bharath Hariharan, and Ser-Nam Lim

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning. InProceedings of the 17th European Conference on Computer Vision, ECCV 2022, pages 709–

    2022

  20. [20]

    Zhanghan Ke, Yuhao Liu, Lei Zhu, Nanxuan Zhao, and Rynson W. H. Lau. Neural preset for color style transfer. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, pages 14173–14182. IEEE, 2023. 1

    2023

  21. [21]

    Uvim: A unified modeling approach for vision with learned guid- ing codes

    Alexander Kolesnikov, Andr ´e Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, and Neil Houlsby. Uvim: A unified modeling approach for vision with learned guid- ing codes. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Information Pro- cessing Systems 2022, NeurIPS 2022, 2022. 2

    2022

  22. [22]

    Viescore: Towards explainable metrics for condi- tional image synthesis evaluation

    Max Ku, Dongfu Jiang, Cong Wei, Xiang Yue, and Wenhu Chen. Viescore: Towards explainable metrics for condi- tional image synthesis evaluation. InProceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), ACL 2024, pages 12268– 12290. Association for Computational Linguistics, 2024. 2, 5

    2024

  23. [23]

    Transient attributes for high-level under- standing and editing of outdoor scenes.ACM Transactions on Graphics (TOG), 33(4):149:1–149:11, 2014

    Pierre-Yves Laffont, Zhile Ren, Xiaofeng Tao, Chao Qian, and James Hays. Transient attributes for high-level under- standing and editing of outdoor scenes.ACM Transactions on Graphics (TOG), 33(4):149:1–149:11, 2014. 6

    2014

  24. [24]

    Yann LeCun, Yoshua Bengio, and Geoffrey E. Hinton. Deep learning.Nature, 521(7553):436–444, 2015. 1

    2015

  25. [25]

    Large lan- guage model-aware in-context learning for code generation

    Jia Li, Chongyang Tao, Jia Li Male, Ge Li, Zhi Jin, Huangzhao Zhang, Zheng Fang, and Fang Liu. Large lan- guage model-aware in-context learning for code generation. ACM Transactions on Software Engineering and Methodol- ogy, 34(7):190:1–190:33, 2025. 1

    2025

  26. [26]

    Unifying image processing as visual prompting question answering

    Yihao Liu, Xiangyu Chen, Xianzheng Ma, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Unifying image processing as visual prompting question answering. InPro- ceedings of the 41st International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 3

    2024

  27. [27]

    UNIFIED-IO: A unified model for vision, language, and multi-modal tasks

    Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mot- taghi, and Aniruddha Kembhavi. UNIFIED-IO: A unified model for vision, language, and multi-modal tasks. InThe Eleventh International Conference on Learning Representa- tions, ICLR 2023. OpenReview.net, 2023. 2

    2023

  28. [28]

    Vi- sion language models are in-context value learners

    Yecheng Jason Ma, Joey Hejna, Chuyuan Fu, Dhruv Shah, Jacky Liang, Zhuo Xu, Sean Kirmani, Peng Xu, Danny Driess, Ted Xiao, Osbert Bastani, Dinesh Jayaraman, Wen- hao Yu, Tingnan Zhang, Dorsa Sadigh, and Fei Xia. Vi- sion language models are in-context value learners. InThe Thirteenth International Conference on Learning Represen- tations, ICLR 2025. OpenR...

    2025

  29. [29]

    Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Re- thinking the role of demonstrations: What makes in-context learning work? InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, pages 11048–11064. Association for Com- putational Linguistics, 2022. 1

    2022

  30. [30]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pages 257–265. IEEE Computer Society, 2017. 6

    2017

  31. [31]

    Grains: Gradient-based attribution for inference- time steering of llms and vlms.CoRR, abs/2507.18043,

    Duy Nguyen, Archiki Prasad, Elias Stengel-Eskin, and Mo- hit Bansal. Grains: Gradient-based attribution for inference- time steering of llms and vlms.CoRR, abs/2507.18043,

    Pith/arXiv arXiv

  32. [32]

    Balasubramanian

    Arghya Pal and Vineeth N. Balasubramanian. Zero-shot task transfer. InIEEE Conference on Computer Vision and Pat- tern Recognition, CVPR 2019, pages 2189–2198, 2019. 1, 3

    2019

  33. [33]

    Khan, and Fahad Shahbaz Khan

    Vaishnav Potlapalli, Syed Waqas Zamir, Salman H. Khan, and Fahad Shahbaz Khan. Promptir: Prompting for all-in- one image restoration. InAdvances in Neural Information Processing Systems 36: Annual Conference on Neural Infor- mation Processing Systems 2023, NeurIPS 2023, 2023. 1, 3

    2023

  34. [34]

    SPIRE: semantic prompt-driven image restoration

    Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbra- cio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi. SPIRE: semantic prompt-driven image restoration. InPro- ceedings of the 18th European Conference on Computer Vi- sion, ECCV 2024, pages 446–464. Springer, 2024. 3

    2024

  35. [35]

    Visual text meets low-level vision: A comprehensive survey on visual text processing.CoRR, abs/2402.03082, 2024

    Yan Shu, Weichao Zeng, Zhenhang Li, Fangmin Zhao, and Yu Zhou. Visual text meets low-level vision: A comprehensive survey on visual text processing.CoRR, abs/2402.03082, 2024. 1

    Pith/arXiv arXiv 2024

  36. [36]

    Suzanna Sia, David Mueller, and Kevin Duh. Where does in-context learning happen in large language models? In Advances in Neural Information Processing Systems 38: An- nual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, 2024. 1

    2024

  37. [37]

    Guibas, Jitendra Malik, and Silvio Savarese

    Trevor Standley, Amir Zamir, Dawn Chen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Which tasks should be learned together in multi-task learning? InPro- ceedings of the 37th International Conference on Machine Learning, ICML 2020, pages 9120–9132. PMLR, 2020. 3

    2020

  38. [38]

    Emu: Generative pretraining in multimodality

    Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. InProceedings of the 12th International Con- ference on Learning Representations, ICLR 2024. OpenRe- view.net, 2024. 2

    2024

  39. [39]

    X-prompt: Towards universal in-context image generation in auto-regressive vision language foundation models.CoRR, abs/2412.01824, 2024

    Zeyi Sun, Ziyang Chu, Pan Zhang, Tong Wu, Xiaoyi Dong, Yuhang Zang, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. X-prompt: Towards universal in-context image generation in auto-regressive vision language foundation models.CoRR, abs/2412.01824, 2024. 1, 3

    Pith/arXiv arXiv 2024

  40. [40]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. InProceedings of the 34th International Conference on Machine Learning, ICML 2017, pages 3319–3328. PMLR, 2017. 2

    2017

  41. [41]

    Chameleon: Mixed-modal early-fusion foundation models.CoRR, abs/2405.09818, 2024

    Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models.CoRR, abs/2405.09818, 2024. 2

    Pith/arXiv arXiv 2024

  42. [42]

    A comprehensive survey of deep learning approaches in image processing.Sensors, 25 (2):531, 2025

    Maria Trigka and Elias Dritsas. A comprehensive survey of deep learning approaches in image processing.Sensors, 25 (2):531, 2025. 1

    2025

  43. [43]

    Renjie Wan, Boxin Shi, Haoliang Li, Yuchen Hong, Ling- Yu Duan, and Alex C. Kot. Benchmarking single-image re- flection removal algorithms.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1424–1441, 2023. 6

    2023

  44. [44]

    Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal

    Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. InIEEE Conference on Com- puter Vision and Pattern Recognition, CVPR 2018, pages 1788–1797. Computer Vision Foundation / IEEE Computer Society, 2018. 6

    2018

  45. [45]

    OFA: unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework

    Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. InInternational Conference on Machine Learn- ing, ICML 2022, pages 23318–23340. PMLR, 2022. 2

    2022

  46. [46]

    Images speak in images: A generalist painter for in-context visual learning

    Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, pages 6830–6839. IEEE, 2023. 1, 2

    2023

  47. [47]

    Large language models are latent variable models: Explaining and finding good demonstra- tions for in-context learning

    Xinyi Wang, Wanrong Zhu, Michael Saxon, Mark Steyvers, and William Yang Wang. Large language models are latent variable models: Explaining and finding good demonstra- tions for in-context learning. InAdvances in Neural Informa- tion Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, 2023. 1

    2023

  48. [48]

    Emu3: Next-token prediction is all you need.CoRR, abs/2409.18869, 2024

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-token prediction is all you need.CoRR,...

    Pith/arXiv arXiv 2024

  49. [49]

    In-context learning unlocked for diffusion models

    Zhendong Wang, Yifan Jiang, Yadong Lu, Yelong Shen, Pengcheng He, Weizhu Chen, Zhangyang (Atlas) Wang, and Mingyuan Zhou. In-context learning unlocked for diffusion models. InAdvances in Neural Information Processing Sys- tems 36: Annual Conference on Neural Information Process- ing Systems 2023, NeurIPS 2023, 2023. 1, 3

    2023

  50. [50]

    Deep retinex decomposition for low-light enhancement

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. In British Machine Vision Conference 2018, BMVC 2018, page

    2018

  51. [51]

    Florence-2: Advancing a unified representation for a vari- ety of vision tasks

    Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a vari- ety of vision tasks. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, pages 4818–

    2024

  52. [52]

    Show-o: One single transformer to unify multimodal understanding and generation

    Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. InProceedings of the 13th International Con- ference on Learning Representations, ICLR 2025. OpenRe- view.net, 2025. 2

    2025

  53. [53]

    Towards efficient and scale-robust ultra-high-definition image demoir ´eing

    Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Jiajun Shen, Jia Li, and Xiaojuan Qi. Towards efficient and scale-robust ultra-high-definition image demoir ´eing. InProceedings of the 17th European Conference on Computer Vision, ECCV 2022, pages 646–662. Springer, 2022. 6

    2022

  54. [54]

    Promptfix: You prompt and we fix the photo

    Yongsheng Yu, Ziyun Zeng, Hang Hua, Jianlong Fu, and Jiebo Luo. Promptfix: You prompt and we fix the photo. In Advances in Neural Information Processing Systems 38: An- nual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, 2024. 3

    2024

  55. [55]

    Zamir, Alexander Sax, William B

    Amir R. Zamir, Alexander Sax, William B. Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, pages 3712–3722. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 3

    2018

  56. [56]

    A comprehensive evaluation of full reference image quality as- sessment algorithms

    Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. A comprehensive evaluation of full reference image quality as- sessment algorithms. In19th IEEE International Conference on Image Processing, ICIP 2012, pages 1477–1480. IEEE,

    2012

  57. [57]

    Perceive-ir: Learning to perceive degrada- tion better for all-in-one image restoration.IEEE Transac- tions on Image Processing, 2025

    Xu Zhang, Jiaqi Ma, Guoli Wang, Qian Zhang, Huan Zhang, and Lefei Zhang. Perceive-ir: Learning to perceive degrada- tion better for all-in-one image restoration.IEEE Transac- tions on Image Processing, 2025. 3

    2025

  58. [58]

    Transfusion: Pre- dict the next token and diffuse images with one multi- modal model

    Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Pre- dict the next token and diffuse images with one multi- modal model. InThe Thirteenth International Conference on Learning Representations, ICLR 2025. OpenReview.net,

    2025

  59. [59]

    Learning to prompt for vision-language models.In- ternational Journal of Computer Vision, 130(9):2337–2348,

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.In- ternational Journal of Computer Vision, 130(9):2337–2348,

  60. [60]

    Seeing the unseen: A fre- quency prompt guided transformer for image restoration

    Shihao Zhou, Jinshan Pan, Jinglei Shi, Duosheng Chen, Lishen Qu, and Jufeng Yang. Seeing the unseen: A fre- quency prompt guided transformer for image restoration. In Proceedings of the 18th European Conference on Computer Vision, ECCV 2024, pages 246–264. Springer, 2024. 3

    2024

  61. [61]

    Visual in-context learning for large vision-language models

    Yucheng Zhou, Xiang Li, Qianning Wang, and Jianbing Shen. Visual in-context learning for large vision-language models. InFindings of the Association for Computational Linguistics, ACL 2024, pages 15890–15902. Association for Computational Linguistics, 2024. 1

    2024

  62. [62]

    Uni-perceiver: Pre- training unified architecture for generic perception for zero- shot and few-shot tasks

    Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre- training unified architecture for generic perception for zero- shot and few-shot tasks. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition, CVPR 2022, pages 16783–16794. IEEE, 2022. 2

    2022

  63. [155]

    BMV A Press, 2018. 6

    2018