Pith. sign in

REVIEW 4 major objections 4 minor 32 references

Dynamic Double Space Tower

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that replacing attention with a four-layer Gestalt 'space tower' lets a 3B multimodal model beat larger models on spatial reasoning benchmarks.

desk verdict Gestalt-inspired spatial tower for VQA: nice idea, but the closure 'prior' is trained against an unnamed ViT detector, the table doesn't show SOTA, and the paper is not reproducible — desk reject. read the letter →

arxiv 2506.11394 v1 pith:GNY65NGO submitted 2025-06-13 cs.CV cs.AI

Weikai Sun , Shijie Song , Han Wang This is my paper
classification cs.CVcs.AI
keywords visualquestionansweringspatialreasoninggestaltprinciplesattentionreplacementmultimodalmodelcausaldynamicdoublespacetower
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 argues that the attention mechanism in multimodal question-answering models limits spatial reasoning because it aligns pixels with text instead of organizing image content into perceptual wholes. The proposed fix is a dynamic double space tower: a four-layer module built on the Gestalt principles of proximity, similarity, closure, and continuity that reorganizes visual features around meaningful units before they are fused with language. On top of this tower the authors add causal and statistical dual-drive reasoning plus text-space-guided cross-causal weights. The resulting 3B-parameter model, July, is reported to reach state-of-the-art scores on spatial-relation benchmarks such as VSI-Bench and CV-Bench, and the tower itself is claimed to be portable to other multimodal models. The practical upshot, if the paper is right, is that spatial ability in VQA can be improved by changing how the model groups visual content, not by scaling parameters.

What carries the argument

The central object is the dynamic double space tower, a four-layer module that replaces the attention mechanism. Each layer implements one Gestalt grouping principle: proximity builds a graph over superpixels and computes attention at node level; similarity clusters regions by visual features and biases attention toward task-relevant clusters; closure detects nearly closed edges, connects them with a virtual bridge, and is trained with an IoU loss against complete-object masks from a ViT-based detection model; continuity guides attention along text-specified paths while separating background from entities. A second named mechanism, the cause-statistics dual-driven framework, adds counterfactual intervention layers to the decoder and shared causal and statistical experts to bias generation toward causal dependencies. Together they convert the model from pixel-text alignment to unit-based perceptual organization.

What would settle it

Remove or corrupt the IoU closure supervision, for example by replacing the ViT mask detector's output with random masks, retrain, and compare VSI-Bench and CV-Bench accuracy; if the spatial-reasoning advantage over attention disappears, the Gestalt closure prior is not what carries the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that the bottleneck in visual question answering is the way attention reads an image: it searches for relationships between pixels, whereas human vision first groups pixels into organized wholes. To exploit that insight, the authors construct a dynamic bidirectional spatial tower with four layers corresponding to proximity, similarity, closure, and continuity. The tower computes attention at the level of superpixels and regions, propagates information on a graph of spatially adjacent regions, closes incomplete object contours through a virtual bridge and an IoU-based supervision against complete-object masks, and follows text-guided paths to preserve continuity. Combined with a causal intervention module and shared causal and statistical experts, this architecture defines the 3B model July, which the paper reports outperforms existing multimodal models of similar or larger size on spatial-relation question-answering benchmarks. The paper also states that the module can be inserted into any multimodal model to improve spatial processing.

Load-bearing premise

The tower's built-in Gestalt closure prior depends on an external ViT-based detection model producing accurate complete-object masks for occluded objects; if that detector is missing or wrong, the claimed spatial-reasoning gain has no support.

Editorial extensions

If this is right

  • If correct, a 3B model can outperform much larger multimodal models on spatial-relation VQA, so spatial reasoning does not require scaling parameters.
  • The tower's portability claim implies that existing attention-based VQA models can be upgraded by swapping in the four-layer module instead of retraining from scratch.
  • On spatial benchmarks such as VSI-Bench, replacing attention with the tower is reported to add about 12 percentage points over a same-size LLaVA-style baseline.
  • Causal intervention layers are said to add less than 5% computation to decoding, so the reasoning gain is claimed to come at small efficiency cost.

Reading between the lines

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

  • Because the tower operates on superpixels and regions rather than raw tokens, part of the reported gain could come from token reduction; ablating token count separately from the Gestalt priors would isolate the mechanism.
  • The same grouping-prior idea could transfer to other spatial tasks involving occlusion and layout, such as visual navigation, robotic manipulation, and 3D scene understanding; these lie outside the paper's experiments.
  • The dependence on an external ViT mask detector means the 'built-in' Gestalt closure prior is partly learned from supervision; a self-supervised closure objective would test whether Gestalt grouping can be acquired without external masks.
  • The causal consistency mechanism may matter most on 'why' and counterfactual questions, which current spatial benchmarks do not isolate; a causal-VQA benchmark would sharpen the comparison.
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.

Referee Report

4 major / 4 minor

Summary. The paper proposes the Dynamic Double Space Tower (DDST), a module intended to replace the standard attention mechanism in multimodal VQA models. The tower is organized around four Gestalt principles (proximity, similarity, closure, continuity) and is combined with a causal dual-driven mechanism using counterfactual intervention and causal role tokens. The authors claim that a 3B-parameter model, called July, achieves state-of-the-art results, especially on spatial-relation VQA, and that the module is portable to any multimodal model. The experimental section reports a single table of results on eight benchmarks and an ablation study described only in prose. The central claims are not supported by the presented evidence: the table shows July trailing several baselines on multiple benchmarks, and the only ablation result is an unverified 12-point gain on VISTBENCH.

Significance. If the proposed architectural ideas worked as claimed, replacing attention with a Gestalt-inspired spatial tower and integrating causal reasoning into a lightweight 3B model would be a meaningful contribution to efficient multimodal reasoning. The paper also attempts to combine multiple recent ideas, such as spatial priors and causal intervention, which is a direction of interest. However, the manuscript in its current form does not substantiate these promises. The only quantitative comparison table is limited and partially contradicts the abstract's state-of-the-art claim; the central closure mechanism relies on an unspecified external detector; and the causal components are described qualitatively without enough detail to be reproduced or tested. The paper does not provide code, model details, or evidence that the module can be inserted into other architectures.

major comments (4)
  1. [Abstract and Table I] The abstract states that the 3B-parameter model July "has achieved state-of-the-art results," but Table I does not support this claim. July trails Qwen2.5VL 3B on vqa-rad (70.9 vs. 78.1) and MMBench (55.6 vs. 56.3), and trails InternVL2 4B on vqa-rad, MMBench, and CLEVR. On path-vqa, the margin over Qwen2.5VL is 0.5 points (79.6 vs. 79.1), yet no error bars, multiple runs, or statistical significance tests are reported. The table also omits many recent state-of-the-art models that the paper claims to outperform, so the central empirical claim is unsupported.
  2. [Section IV, Ablation experiments] The only quantitative evidence for the effectiveness of the dynamic double space tower is the sentence that embedding the tower instead of attention gives "12 percentage point higher accuracy on VISTBENCH." No ablation table, no standard deviations, no description of the VISTBENCH subset, and no results for the other proposed components are provided. This single assertion cannot be verified and is not sufficient to support the portability and improvement claims made in the abstract and introduction.
  3. [Section IV, Implementation details] The closure layer, one of the four Gestalt principles on which the tower is built, is trained by computing IoU between the "space tower prior map" and the "mask of the complete object" obtained from an unnamed "vit-based visual detection model." The paper does not specify which detector is used, whether it outputs amodal (occlusion-completed) masks, or how the continuous prior map is binarized for the IoU calculation. Most off-the-shelf ViT segmenters produce visible-region masks; if such masks were used, the loss does not learn closure at all. As reported, the closure mechanism is unfalsifiable and cannot be distinguished from simply fitting another network's outputs.
  4. [Section III.C and III.D] The causal dual-driven mechanism, including the counterfactual intervention module, the causal expert system, the [CAUSE]/[EFFECT] tokens, and the causal role vectors E_cause and E_effect, is described only in qualitative terms. No equations are given for the causal loss L_causal-cls, no dimensions or initialization are specified for the role vectors, no hyperparameters are reported, and the claim that the intervention layers increase computation by "less than 5%" is not backed by any measurement. This makes the causal contribution impossible to reproduce and evaluate.
minor comments (4)
  1. [Section III.A, Figure 4 caption] The caption reads "Continuty architecture" and should be "Continuity architecture."
  2. [Section IV] The text refers to "Figure 1" when presenting results, but the results are shown in Table I; the figure-reference needs to be corrected.
  3. [Section V, Conclusions] The final paragraph discusses the "MSPG-SEN architecture" and image generation, which is unrelated to DDST and appears to be a copy-paste artifact; it should be removed or rewritten to match the paper's content.
  4. [Throughout] The paper contains numerous grammatical errors and informal phrasings (e.g., "Change from seeing images to perceiving and organizing image content"), and many citations in the introduction are placed in clumps without clear connection to the statements they support. A thorough editorial revision is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DDST module is trained with supervised losses, including an IoU loss against a ViT detector, but the paper's central SOTA claim rests on independent external benchmarks and is not forced by construction.

full rationale

The paper's derivation chain is not circular. The four Gestalt layers are engineering choices (proximity/similarity via superpixels and features, closure via edge bridging and an IoU loss, continuity via text-space-guided weights), not results derived from first principles; the paper then measures their effect on external benchmarks (MMBench, POPE, CV-Bench, VSI-Bench, CLEVR, etc.). The one genuinely load-bearing supervision, the closure prior, is trained by IoU against masks from a 'vit-based visual detection model' (Section IV), meaning the 'Gestalt closure prior' is fitted to another network's output rather than coming from a closed-form Gestalt derivation. This is an unvalidated external dependency and a correctness/support risk, but it is not a circularity: the VQA/spatial-reasoning scores could in principle differ from that detector's masks, and the SOTA claim is an empirical comparison against external models rather than a quantity that equals the training loss by construction. No self-citations, uniqueness theorems, or ansatz-smuggling citations are load-bearing, and no prediction is the by-construction re-statement of a fitted parameter. Minor issues such as 'prior' being learned, unspecified detector details, and possible benchmark overlap are better categorized as reproducibility/correctness concerns, not as circular reasoning.

Assumptions & free parameters 4 free parameters · 5 assumptions · 5 invented entities

The paper presents no formal derivation. The 'structural prior' it claims is trained against an external detector and VQA labels, so the free parameters and invented modules carry the entire explanatory load. All empirical claims depend on undisclosed training details.

free parameters (4)
  • Number of tower layers = 4
    Hand-chosen to match four Gestalt principles; no ablation shows this is optimal.
  • Causal role vectors E_cause, E_effect
    No values or training rule given; assignment relies on dependency parsing or pretrained dictionaries.
  • Closure IoU loss weight
    IoU term is added to the loss with no coefficient specified, making it an ad hoc fitted knob.
  • Causal intervention depth = Last two decoder layers
    Hand-chosen; no experiment varies this choice.
assumptions (5)
  • domain assumption Gestalt grouping principles are an effective inductive bias for VQA spatial reasoning.
    The entire tower design in Section III.A is justified by these principles without empirical support specific to VQA.
  • ad hoc to paper A ViT-based object detector supplies accurate complete-object masks for the closure loss.
    Section IV relies on this detector for the IoU loss, but no detector is named or its accuracy analyzed.
  • ad hoc to paper Causal trigger words can be reliably detected with spaCy/NLTK and role tags improve reasoning.
    Section III.D assumes this machinery yields a correct cause/effect decomposition; no evidence is given.
  • domain assumption CNN features can represent graph structure for proximity propagation.
    Section III.B states CNNs disseminate information on graph structure without formal justification.
  • ad hoc to paper The three-stage training procedure is sufficient to learn the claimed spatial and causal abilities.
    Stages are listed in Section IV, but no ablation isolates their contribution.
invented entities (5)
  • Dynamic double space tower
    purpose: Replace attention to model spatial relationships
    No formal definition, code, or weights are provided, so there is no external falsifiable handle.
  • Virtual bridge
    purpose: Connect nearly closed edges to complete object outlines
    Mentioned only in the closure module description; no implementation or measured effect.
  • Causal expert system
    purpose: Combine causal and statistical reasoning in a lightweight parameter-shared mixture
    Described as sharing a bottom network; no routing details, capacity, or isolated results.
  • [CAUSE]/[EFFECT] tokens
    purpose: Mark causal intent in questions
    Text-only markers; their effect is not isolated in any experiment.
  • Causal role vectors E_cause, E_effect
    purpose: Encode cause/effect roles in position encoding
    Vectors are never specified or released.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic Double Space Tower." pith.science (2026). https://pith.science/paper/GNY65NGO

@misc{pith2026250611394,
  author       = {Pith},
  title        = {Pith review of: Dynamic Double Space Tower},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GNY65NGO}},
  note         = {Machine review of arXiv:2506.11394}
}
read the original abstract

The Visual Question Answering (VQA) task requires the simultaneous understanding of image content and question semantics. However, existing methods often have difficulty handling complex reasoning scenarios due to insufficient cross-modal interaction and capturing the entity spatial relationships in the image.\cite{huang2023adaptive}\cite{liu2021comparing}\cite{guibas2021adaptive}\cite{zhang2022vsa}We studied a brand-new approach to replace the attention mechanism in order to enhance the reasoning ability of the model and its understanding of spatial relationships.Specifically, we propose a dynamic bidirectional spatial tower, which is divided into four layers to observe the image according to the principle of human gestalt vision. This naturally provides a powerful structural prior for the spatial organization between entities, enabling the model to no longer blindly search for relationships between pixels but make judgments based on more meaningful perceptual units. Change from "seeing images" to "perceiving and organizing image content".A large number of experiments have shown that our module can be used in any other multimodal model and achieve advanced results, demonstrating its potential in spatial relationship processing.Meanwhile, the multimodal visual question-answering model July trained by our method has achieved state-of-the-art results with only 3B parameters, especially on the question-answering dataset of spatial relations.

Figures

Figures reproduced from arXiv: 2506.11394 by the authors.

Figure 2
Figure 2. Similarity architecture Thirdly, for closure, it refers to the fact that the brain will automatically fill in incomplete graphics to make them a complete whole. We assume that when the model recognizes certain parts of an object, it should be able to infer the outline or range of the complete object and direct attention to this ”closed” area, even if some parts are occluded or unclear. We combine classic computer vi… view at source ↗
Figure 1
Figure 1. Proximity architecture Secondly, regarding similarity, it refers to the fact that elements with similar attributes will be automatically grouped into the same group or whole by the brain, even if they are not spatially close. This similarity can be reflected in multiple visual dimensions. The brain simplifies information processing and enhances perceptual efficiency by capturing these commonalities. Based on this, w… view at source ↗
Figure 3
Figure 3. Closure architecture Fourth, regarding continuity, it refers to the visual tendency to perceive continuous lines or shapes rather than broken fragments. Let’s make the model for something like ”along...” ”Through...” When it comes to the relationship of paths or directions, we guide the continuous module to gradually move in the direction of the path through the weights guided by the text space. ”Remember” the spati… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Continuty architecture B. Model architecture Our model consists of three parts. The first part is the encoding part, the second part is the embedding and fusion, and the third part is the decoding and generation part. For the encoding part, we choose to improve on the …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 18 canonical work pages

  1. [1]

    Vlmo: Unified vision-language pre-training with mixture-of-modality- experts.Advances in Neural Information Processing Systems, 35:32897– 32912, 2022

    Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mo- hammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality- experts.Advances in Neural Information Processing Systems, 35:32897– 32912, 2022

    work page 2022

  2. [2]

    Automatic image processing algorithm for light en- vironment optimization based on multimodal neural network model

    Mujun Chen. Automatic image processing algorithm for light en- vironment optimization based on multimodal neural network model. Computational Intelligence and Neuroscience, 2022(1):5156532, 2022

    work page 2022

  3. [3]

    Adaptive fourier neural operators: Efficient token mixers for transformers.arXiv preprint arXiv:2111.13587, 2021

    John Guibas, Morteza Mardani, Zongyi Li, Andrew Tao, Anima Anand- kumar, and Bryan Catanzaro. Adaptive fourier neural operators: Efficient token mixers for transformers.arXiv preprint arXiv:2111.13587, 2021. DYNAMIC DOUBLE SPACE TOWER 7

    arXiv 2021

  4. [4]

    Crd-cgan: Category- consistent and relativistic constraints for diverse text-to-image genera- tion.Frontiers of Computer Science, 18(1):181304, 2024

    Tao Hu, Chengjiang Long, and Chunxia Xiao. Crd-cgan: Category- consistent and relativistic constraints for diverse text-to-image genera- tion.Frontiers of Computer Science, 18(1):181304, 2024

    work page 2024

  5. [5]

    Language is not all you need: Aligning perception with language models.Advances in Neural Information Processing Systems, 36:72096–72109, 2023

    Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, et al. Language is not all you need: Aligning perception with language models.Advances in Neural Information Processing Systems, 36:72096–72109, 2023

    2023

  6. [6]

    Adaptive frequency filters as efficient global token mixers

    Zhipeng Huang, Zhizheng Zhang, Cuiling Lan, Zheng-Jun Zha, Yan Lu, and Baining Guo. Adaptive frequency filters as efficient global token mixers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 6049–6059, 2023

    work page 2023

  7. [7]

    Fast and efficient image generation using variational autoencoders and k-nearest neighbor oversampling approach.IEEE Access, 11:28416–28426, 2023

    Ashhadul Islam and Samir Brahim Belhaouari. Fast and efficient image generation using variational autoencoders and k-nearest neighbor oversampling approach.IEEE Access, 11:28416–28426, 2023

    work page 2023

  8. [8]

    An underwater image enhancement method for a preprocessing framework based on generative adversarial network

    Xiao Jiang, Haibin Yu, Yaxin Zhang, Mian Pan, Zhu Li, Jingbiao Liu, and Shuaishuai Lv. An underwater image enhancement method for a preprocessing framework based on generative adversarial network. Sensors, 23(13):5774, 2023

    2023

Show all 32 references
  1. [9]

    Vilt: Vision-and-language transformer without convolution or region supervision

    Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. InInternational conference on machine learning, pages 5583–5594. PMLR, 2021

    2021

  2. [10]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  3. [11]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InInternational conference on machine learning, pages 12888–12900. PMLR, 2022

    2022

  4. [12]

    Align before fuse: Vision and language representation learning with momentum distillation.Ad- vances in neural information processing systems, 34:9694–9705, 2021

    Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation.Ad- vances in neural information processing systems, 34:9694–9705, 2021

    2021

  5. [13]

    Vision- language foundation models as effective robot imitators.arXiv preprint arXiv:2311.01378, 2023

    Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision- language foundation models as effective robot imitators.arXiv preprint arXiv:2311.01378, 2023

    2023 arXiv

  6. [14]

    Scaling language-image pre-training via masking

    Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23390–23400, 2023

    2023

  7. [15]

    Novel creation method of feature graphics for image generation based on deep learning algorithms.Mathematics, 11(7):1644, 2023

    Ying Li and Ye Tang. Novel creation method of feature graphics for image generation based on deep learning algorithms.Mathematics, 11(7):1644, 2023

    2023

  8. [16]

    Wei Liu, Jie-Lin Qiu, Wei-Long Zheng, and Bao-Liang Lu. Comparing recognition performance and robustness of multimodal deep learning models for multimodal emotion recognition.IEEE Transactions on Cognitive and Developmental Systems, 14(2):715–729, 2021

    2021

  9. [17]

    Cogan: Cooperatively trained conditional and unconditional gan for person image generation.IET Image Processing, 17(10):2949–2957, 2023

    Yang Liu, Hao Sheng, Shuai Wang, Yubin Wu, and Zhang Xiong. Cogan: Cooperatively trained conditional and unconditional gan for person image generation.IET Image Processing, 17(10):2949–2957, 2023

    2023

  10. [18]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

    2021

  11. [19]

    Explor- ing the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Explor- ing the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1–67, 2020

    2020

  12. [20]

    Generative multimodal models are in-context learners

    Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages ...

    2024

  13. [21]

    Attentional generative adversarial networks with representativeness and diversity for generating text to realistic image

    Anjie Tian and Lu Lu. Attentional generative adversarial networks with representativeness and diversity for generating text to realistic image. IEEE access, 8:9587–9596, 2020

    2020

  14. [22]

    Improving the quality of image generation in art with top-k training and cyclic generative methods.Scientific Reports, 13(1):17764, 2023

    Laura Vela, F ´elix Fuentes-Hurtado, and Adri ´an Colomer. Improving the quality of image generation in art with top-k training and cyclic generative methods.Scientific Reports, 13(1):17764, 2023

    2023

  15. [23]

    Image generation and recognition technology based on attention residual gan.IEEE Access, 11:61855–61865, 2023

    Huazhe Wang and Li Ma. Image generation and recognition technology based on attention residual gan.IEEE Access, 11:61855–61865, 2023

    2023

  16. [24]

    Image as a foreign language: Beit pretraining for all vision and vision-language tasks.arXiv preprint arXiv:2208.10442, 2022

    Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks.arXiv preprint arXiv:2208.10442, 2022

    2022 arXiv

  17. [25]

    Rca-gan: An improved image denoising algorithm based on generative adversarial networks.Electronics, 12(22):4595, 2023

    Yuming Wang, Shuaili Luo, Liyun Ma, and Min Huang. Rca-gan: An improved image denoising algorithm based on generative adversarial networks.Electronics, 12(22):4595, 2023

    2023

  18. [26]

    Genartist: Multimodal llm as an agent for unified image generation and editing

    Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024

    2024 arXiv

  19. [27]

    Lpgan: A lbp-based proportional input generative adversarial network for image fusion.Remote Sensing, 15(9):2440, 2023

    Dongxu Yang, Yongbin Zheng, Wanying Xu, Peng Sun, and Di Zhu. Lpgan: A lbp-based proportional input generative adversarial network for image fusion.Remote Sensing, 15(9):2440, 2023

    2023

  20. [28]

    Vsa: Learning varied-size window attention in vision transformers

    Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vsa: Learning varied-size window attention in vision transformers. InEuropean conference on computer vision, pages 466–483. Springer, 2022

    2022

  21. [29]

    Joint generative image deblurring aided by edge attention prior and dynamic kernel selection.Wireless Communications and Mobile Computing, 2021(1):1391801, 2021

    Zhichao Zhang, Hui Chen, Xiaoqing Yin, and Jinsheng Deng. Joint generative image deblurring aided by edge attention prior and dynamic kernel selection.Wireless Communications and Mobile Computing, 2021(1):1391801, 2021

    2021

  22. [30]

    Cyclic generative attention- adversarial network for low-light image enhancement.Sensors, 23(15):6990, 2023

    Tong Zhen, Daxin Peng, and Zhihui Li. Cyclic generative attention- adversarial network for low-light image enhancement.Sensors, 23(15):6990, 2023

    2023

  23. [31]

    Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration

    Shihao Zhou, Duosheng Chen, Jinshan Pan, Jinglei Shi, and Jufeng Yang. Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2952– 2963, 2024

    2024

  24. [32]

    Wdig: a wavelet domain image generation framework based on frequency domain op- timization.EURASIP Journal on Advances in Signal Processing, 2023(1):66, 2023

    Qing Zhu, Xiumei Li, Junmei Sun, and Huang Bai. Wdig: a wavelet domain image generation framework based on frequency domain op- timization.EURASIP Journal on Advances in Signal Processing, 2023(1):66, 2023. Sun WeikaiHe is currently studying for a Bache- lor’s degree in Comput...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.