Pith. sign in

REVIEW 4 major objections 6 minor 50 references

ArchiLense: A Framework for Quantitative Analysis of Architectural Styles Based on Vision Large Language Models

T0 review · 4 major / 6 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 ArchiLense claims that a vision-language-model pipeline can generate, rank, and validate textual descriptions of architectural styles directly from images, reaching 84.5% expert matching accuracy and a 4.62 mean expert score.

desk verdict The paper's headline metrics don't hold up, but the dataset idea and pipeline are worth a serious revision. read the letter →

arxiv 2506.07739 v3 pith:TR3CFBGM submitted 2025-06-09 cs.CV cs.AI

Jing Zhong , Jun Yin , Peilin Li , Pengyu Zeng , Miao Zang , Ran Luo , Shuai Lu This is my paper
classification cs.CVcs.AI
keywords architecturalstyleanalysisvisionlanguagemodelsCLIPdescriptiongenerationArchDiffBenchtext-to-imageexpertevaluationquantitativecomparison
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

The paper tries to establish that architectural style analysis, traditionally done through subjective expert reading and historical review, can be made quantitative and repeatable with vision-language models. It introduces ArchiLense, which automatically produces textual descriptions of what makes one group of buildings stylistically different from another, then ranks those descriptions by how well they separate the groups. The framework is built on a new dataset, ArchDiffBench, of 1,765 images of works by ten architects, organized into 81 pairwise comparisons. The authors report that the ranked descriptions separate the architect groups with statistically significant t-tests and that human experts match the generated style descriptions to the right architects 84.5% of the time, with a mean consistency score of 4.62 out of 5. The abstract additionally reports a 92.4% consistency rate with expert annotations, a figure the full-text experiments do not report.

What carries the argument

The load-bearing mechanism is the style-discrimination score $S_y$ for a candidate style description $y$: $$S_y = \sum_{x \in D_A} \$\theta$(x, y) - \sum_{x \in D_B} \$\theta$(x, y),$$ where $\theta(x, y)$ is the cosine similarity between a CLIP ViT-G/14 embedding of image $x$ and the embedding of description $y$: $$\$\theta$(x, y) = \frac{e_x \cdot e_y}{\|e_x\| \|e_y\|}.$$ Descriptions are ranked by the AUROC of $\theta$ over the two groups, and a t-test at the 0.05 level removes statistically insignificant descriptions. This score is what converts qualitative textual claims about style into a quantitative, rankable, and testable form.

What would settle it

Hold out one or more architect groups entirely, generate style descriptions only from the remaining groups, then apply the ranking to the held-out groups' images and compare the CLIP AUROC against chance. If the AUROC on held-out groups is near 0.5, the descriptions do not generalize and the central claim fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a two-stage vision-language pipeline can turn architectural images into discriminative style descriptions without needing a supervised classifier trained on style labels. The Style Extractor compares text descriptions of individual images across two architect groups to generate candidate style phrases, and the Style Evaluator scores each phrase by the difference in CLIP embedding similarity between the two groups. The authors find that this text-based extraction beats visual-only alternatives, that the CLIP-based evaluator beats VQA- and caption-based matching, and that the resulting descriptions align with expert design knowledge, as shown by word clouds and a pairwise style-similarity matrix among architects.

Load-bearing premise

The whole framework collapses if CLIP cosine similarity between an image and a style description is not a faithful measure of whether the building actually exhibits that style; if it is not, the ranked descriptions may simply separate images on irrelevant visual features.

Editorial extensions

If this is right

  • Style descriptions for any pair of architect groups can be generated automatically, making cross-region and cross-period comparisons reproducible rather than dependent on a single expert's eye.
  • The ranked descriptions can be used to synthesize representative architectural images through text-to-image generation, turning style analysis into a design-generation brief.
  • The pairwise similarity matrix among architects offers a quantitative map of stylistic kinship, such as the reported 0.995 similarity among Meier, Mies, and Kahn.
  • The same extractor-evaluator design could be applied to other visual domains where qualitative expert judgments are the current standard, such as furniture or urban design styles.
  • The pipeline requires no predefined style labels, so it can discover stylistic groupings that were not specified in advance.

Reading between the lines

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

  • Because the descriptions are generated and scored on the same ten architect groups, the reported AUROC and significance values likely overstate how well the framework would separate buildings it has never seen; a held-out architect split would be a stricter test.
  • If the method is correct, the 0.995 Meier/Mies/Kahn link becomes a falsifiable claim about modernist kinship that could be checked against independent art-historical scholarship.
  • A direct extension would replace CLIP with a domain-specific architectural embedding trained on expert pairwise judgments, allowing the framework to be recalibrated to human style perception rather than relying on generic image-text similarity.
  • The abstract's 92.4% consistency figure should be reconciled with the full-text 84.5% matching accuracy before the headline numbers are cited in later work.
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 / 6 minor

Summary. The manuscript introduces ArchiLense, a VLLM-based framework for quantitative architectural style analysis, together with ArchDiffBench, a dataset of 1,765 images grouped into 10 architect categories. The framework generates stylistic descriptions via text-based comparison of image captions (Section 4.1), ranks them using a CLIP cosine-similarity discriminator and t-tests (Section 4.2 and Section 5.2.1), and validates the outputs through text-to-image generation followed by expert evaluation (Section 5.3). The abstract claims a 92.4% consistency rate with expert annotations and 84.5% classification accuracy, and states that the framework 'transcends the subjectivity inherent in traditional analyses.'

Significance. The paper targets an underexplored and culturally important problem: making architectural style comparisons repeatable and quantitative. The idea of generating linguistically articulated style descriptions and ranking them with a discriminative score is attractive, and the authors are transparent about using standard off-the-shelf models such as GPT-4, CLIP, and BLIP-2. However, the reported experiments do not support the headline claims. The 92.4% consistency rate never appears in the body, and the 84.5% figure is a human matching accuracy on generated images, not an automatic classification accuracy of ArchiLense on original images. In addition, the evaluation is circular: descriptions are generated from and tested on the same image groups, so the reported discriminative power does not indicate generalization to unseen buildings. These issues currently preclude acceptance.

major comments (4)
  1. [Abstract and Section 5.3] The abstract's claim of '92.4% consistency rate with expert annotations and 84.5% classification accuracy' is not supported by the full text. The 92.4% figure does not appear anywhere in the body, and the 84.5% figure in Section 5.3 is the accuracy with which human participants matched text-generated images to one of 10 architects, not an automatic classification accuracy of ArchiLense on ArchDiffBench images. No classification experiment on the original dataset is reported in Section 5, so the central empirical claim is unsupported.
  2. [Sections 4.1, 4.2, and 5.2.1] The style descriptions are generated from randomly sampled images of the architect groups (Section 4.1) and then evaluated on the same image groups: Section 5.2.1 reports t-tests on '270 style descriptions from 9 architect groups' using the same groups from which the descriptions were derived. The Sy ranking in Section 4.2 therefore selects descriptions that best separate the training images, and the t-test results do not test generalization to new buildings or new architects. No held-out split, cross-validation, or architect-level evaluation is provided.
  3. [Section 4.2, Eq. (1)-(2)] The entire discriminator pipeline rests on the assumption that CLIP ViT-G/14 cosine similarity between an image and a textual description is a valid proxy for architectural style match. This equivalence is asserted without calibration against human judgment on the original images, and no AUROC values, confidence intervals, or comparisons with conventional CNN-based style classifiers (such as the channel-spatial attention model cited as reference [26]) are reported. Without such validation, the descriptions selected as 'discriminative' may separate images on spurious visual correlates rather than on architectural style.
  4. [Section 6] The authors explicitly acknowledge that the dataset is biased toward modern Western styles and that spatial authenticity analysis is lacking. Given this self-stated limitation and the absence of a non-circular evaluation, the abstract's claim that ArchiLense 'transcends the subjectivity inherent in traditional analyses' and offers objective cross-cultural comparison is premature and not established by the reported experiments.
minor comments (6)
  1. [Section 3] The dataset is not released, and the paper does not provide a detailed breakdown of image counts per architect, resolution, or selection criteria; this limits reproducibility and makes it impossible to verify the claim of '1,765 high-quality images.'
  2. [Section 3 and Figure 4] The paper states '81 paired subsets' in Section 3 and '81 pairwise comparisons' in Figure 4, but there are only 45 pairwise combinations of 10 architects; the discrepancy needs clarification.
  3. [Section 5.3] The expert evaluation lacks methodological detail: it does not report the number of generated images per architect, the instructions given to participants, the inter-rater agreement among the 20 PhD students and 5 senior architects, or how the 'matching accuracy' is computed across raters and items.
  4. [Keywords and Introduction] The keyword list ('Architectural Floor Plan Generation; Transformer Model; Convolutional Neural Networks; Hybrid Deep Learning Models') does not correspond to the paper's content and appears to be from a different manuscript; also, the first sentence of Section 1 uses 'we proposes' and should be corrected.
  5. [References] Reference formatting is inconsistent: several arXiv preprints are cited without venues, one reference contains a garbled author list, and the in-text citation of 'Visdiff [22]' points to a paper about Midjourney architectural virtual labs rather than the Visdiff work described in the text.
  6. [Section 5.2.3] The similarity values in Figure 8 (e.g., 0.995 between Meier, Mies, and Louis Kahn) are mentioned without defining the similarity measure or the aggregation procedure used to produce the matrix, which makes the quantitative claims difficult to interpret.

Circularity Check

1 steps flagged · score 6.0 of 10

The 'discriminative effectiveness' validation is in-sample: descriptions are generated from, AUROC-ranked on, and t-tested on the same architect groups; the abstract's headline metrics (92.4% consistency, 84.5% 'classification accuracy') are not reproduced in the full text.

  1. fitted input called prediction [Sections 4.1-4.2 and 5.2.1 (Sy formula, AUROC ranking, t-test on 270 descriptions)]
    "The Style Extractor aims to identify distinct stylistic features between two architectural groups DA and DB ... It calculates the differentiation fraction Sy of the style description y ... Sy = Σ(x∈DA) θ(x,y) − Σ(x∈DB) θ(x,y) ... AUROC is used to assess the discriminative power of style descriptions between DA and DB. ... An experiment using 270 style descriptions from 9 architect groups ... followed by t-tests. ... over 80% of the descriptions achieved statistical significance (p < 0.05), confirming their effectiveness in differentiating architectural styles."

    The descriptions are generated from the very groups they are validated on. Section 4.1 describes a Style Extractor that produces stylistic difference descriptions from images of two groups DA and DB; Section 4.2 then scores each description with Sy = Σ(x∈DA)θ(x,y) − Σ(x∈DB)θ(x,y), ranks it by AUROC 'between DA and DB', and t-tests the θ distributions of those same groups. The paper describes no held-out set or split anywhere.

full rationale

The clearest circular step is the in-sample validation of the style descriptions (Sections 4.1-4.2 and 5.2.1): descriptions are extracted from pairs of architect groups and then scored, AUROC-ranked, and t-tested on those very groups, so the 'over 80% statistically significant' claim is a within-fit statistic rather than a test on held-out data. This is a fitted-input-called-prediction pattern: the extraction plus the ranking/filtering procedure is the fit, and the reported discriminative effectiveness is measured on the fitting data, so the central quantitative claim of Section 5.2 partially reduces to the selection procedure. I do not count the following as circular: (a) the numerous self-citations (refs 3-5, 8, 10-19, 27, 29-42, 46-50) are not load-bearing, since the framework is inspired by external work (Visdiff, ref 22) and no prior result or uniqueness theorem by the same authors is invoked to force the framework's choices; (b) the abstract's '92.4% consistency rate' never appears in the full text and the '84.5% classification accuracy' is actually human matching accuracy on text-generated images (Section 5.3), not ArchiLense classification of real ArchDiffBench images, so these are claim-support gaps and mislabeled metrics, which are correctness risks rather than circular reductions; (c) the CLIP cosine-similarity proxy for stylistic match (Section 4.2) is an uncalibrated assumption and a validity risk, not a circularity. The expert study is partly self-referential (generated images are rated against the descriptions they were generated from), but it does invoke external human judgment and retains some independent content. Overall score 6: one central 'prediction' (discriminative effectiveness) partially reduces by construction, while the rest of the framework still contains independent, externally grounded components.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper introduces no numerical free parameters, but it rests on several domain assumptions about what VLMs and CLIP scores mean for architectural style. The most expensive assumption is the closed-loop evaluation: descriptions are generated and scored on the same images. The dataset's representativeness is also a limitation, acknowledged by the authors.

assumptions (5)
  • domain assumption Text-based descriptions generated by VLMs capture architect-specific stylistic features better than image-grid or embedding-difference approaches (Section 4.1).
    The paper chooses text-based comparison as the primary extraction strategy based on 'experimental results' but gives no quantitative comparison or independent validation of this claim.
  • domain assumption CLIP ViT-G/14 cosine similarity between an image and a style description is a valid measure of stylistic match (Section 4.2).
    The entire Sy ranking and AUROC evaluation depend on this equivalence, which is asserted without calibration against human judgment.
  • domain assumption The 1,765 images from ArchDaily and gooood, labeled by architect identity, are representative of the architects' styles (Section 3).
    The paper acknowledges a bias toward modern Western styles, so the representativeness for cross-cultural claims is weak.
  • domain assumption Architecture PhD students and senior architects provide reliable ground truth for style consistency (Section 5.3).
    Expert evaluation is the only external validation, but the number of raters is small and inter-rater agreement is not reported.
  • standard math Statistical significance at p<0.05 in a t-test implies practical discriminative validity of a style description (Section 5.2.1).
    Standard methodology, but the test is applied to descriptions and images from the same source, so significant results may reflect overfitting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ArchiLense: A Framework for Quantitative Analysis of Architectural Styles Based on Vision Large Language Models." pith.science (2026). https://pith.science/paper/TR3CFBGM

@misc{pith2026250607739,
  author       = {Pith},
  title        = {Pith review of: ArchiLense: A Framework for Quantitative Analysis of Architectural Styles Based on Vision Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TR3CFBGM}},
  note         = {Machine review of arXiv:2506.07739}
}
read the original abstract

Architectural cultures across regions are characterized by stylistic diversity, shaped by historical, social, and technological contexts in addition to geograph-ical conditions. Understanding architectural styles requires the ability to describe and analyze the stylistic features of different architects from various regions through visual observations of architectural imagery. However, traditional studies of architectural culture have largely relied on subjective expert interpretations and historical literature reviews, often suffering from regional biases and limited ex-planatory scope. To address these challenges, this study proposes three core contributions: (1) We construct a professional architectural style dataset named ArchDiffBench, which comprises 1,765 high-quality architectural images and their corresponding style annotations, collected from different regions and historical periods. (2) We propose ArchiLense, an analytical framework grounded in Vision-Language Models and constructed using the ArchDiffBench dataset. By integrating ad-vanced computer vision techniques, deep learning, and machine learning algo-rithms, ArchiLense enables automatic recognition, comparison, and precise classi-fication of architectural imagery, producing descriptive language outputs that ar-ticulate stylistic differences. (3) Extensive evaluations show that ArchiLense achieves strong performance in architectural style recognition, with a 92.4% con-sistency rate with expert annotations and 84.5% classification accuracy, effec-tively capturing stylistic distinctions across images. The proposed approach transcends the subjectivity inherent in traditional analyses and offers a more objective and accurate perspective for comparative studies of architectural culture.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages

  1. [22]

    Sukkar, A.W., Fareed, M.W., Yahia, M.W., Abdalla, S.B., Ibrahim, I., & Senjab, K.A.K. (2024). Analytical Evaluation of Midjourney Architectural Virtual Lab: Defining Major Cur- rent Limits in AI -Generated Representations of Islamic Architectural Heritage. Buildings, 14(3), 786. DOI: 10.3390/buildings14030786

    work page doi:10.3390/buildings14030786 2024

  2. [26]

    Wang, B., Zhang, S., Zhang, J., & Cai, Z. (2023). Architectural style classification based on CNN and channel–spatial attention. Signal, Image and Video Processing, 17(1), 99-107

    work page 2023

  3. [1]

    Canizaro, V. B. (2007). Critical regionalism: Architecture and identity in a globalized world. Princeton Architectural Press

    work page 2007

  4. [2]

    Chen, F., Mai, M., Huang, X., & Li, Y. (2024). Enhancing the Sustainability of AI Technolo- gy in Architectural Design: Improving the Matching Accuracy of Chinese -Style Buildings. Sustainability, 16(19), 8414. DOI: 10.3390/su16198414

    work page doi:10.3390/su16198414 2024

  5. [3]

    Gao, C., Sun, Y., Yang, K., Xie, Y., Liu, J., & Liu, Y. (2025). Thought -augmented policy op- timization: Bridging external guidance and internal capabilities. arXiv preprint arXiv:2505.15692. https://arxiv.org/abs/2505.15692

    arXiv 2025

  6. [4]

    Gao, W., Lu, S., Zhang, X., He, Q., Huang, W., & Lin, B. (2023). Impact of 3D modeling be- havior patterns on the creativity of sustainable building design through process min- ing. Automation in Construction, 150, 104804

    work page 2023

  7. [5]

    Huang, Y., Zeng, T., Jia, M., Yang, J., Xu, W., & Lu, S. (2025). Fusing Transformer and dif- fusion for high -resolution prediction of daylight illuminance and glare based on sparse ceil- ing-mounted input. Building and Environment, 267, 112163

    work page 2025

  8. [6]

    & Wang, X

    He, Y., Li, S., Li, K., Wang, J., Li, B., Shi, T., ... & Wang, X. (2025). Enhancing Low -Cost Video Editing with Lightweight Adaptors and Temporal -Aware Inversion. arXiv preprint arXiv:2501.04606

    arXiv 2025

Show all 50 references
  1. [7]

    He, Y., Wang, J., Wu, J., Zhu, W., & Liu, Z. (2025). Boosting multimodal reasoning with MCTS-automated structured thinking. arXiv preprint arXiv:2502.02339. https://arxiv.org/abs/2502.02339

    2025

  2. [8]

    How real -time energy feedback influences energy-efficiency and aes- thetics of architecture design and judgment of architects: two design experiments

    Jia, Z., Lu, S., & Yin, J. How real -time energy feedback influences energy-efficiency and aes- thetics of architecture design and judgment of architects: two design experiments

  3. [9]

    Lefaivre, L., & Tzonis, A. (2020). Architecture of Regionalism in the Age of Globalization: Peaks and Valleys in the Flat World (2nd ed.). Routledge. https://doi.org/10.4324/9780367281182

    2020 doi

  4. [10]

    Li, J., Lu, S., Wang, W., Huang, J., Chen, X., & Wang, J. (2018). Design and Climate‐ Responsiveness Performance Evaluation of an Integrated Envelope for Modular Prefabricated Buildings. Advances in Materials Science and Engineering, 2018(1), 8082368

    2018

  5. [11]

    Li, J., Lu, S., Wang, Q., Tian, S., & Jin, Y. (2019). Study of passive adjustment performance of tubular space in subway station building complexes. Applied Sciences, 9(5), 834

    2019

  6. [12]

    Li, J., Song, Y., Lv, S., & Wang, Q. (2015). Impact evaluation of the indoor environmental performance of animate spaces in buildings. Building and environment, 94, 353-370

    2015

  7. [13]

    Li, J., Lu, S., & Wang, Q. (2018). Graphical visualisation assist analysis of indoor environ- mental performance: Impact of atrium spaces on public buildings in cold climates. Indoor and Built Environment, 27(3), 331-347. 10

    2018

  8. [14]

    Outdoor space design and its effect on mental work performance in a subtropical climate

    Luo Y, Ma X, Li J, Wang C, de Dear R, Lu S. Outdoor space design and its effect on mental work performance in a subtropical climate. Building and Environment. 2025 Feb 15;270:112470

    2025

  9. [15]

    Lu, S., Yan, X., Li, J., & Xu, W. (2016). The influence of shape design on the acoustic per- formance of concert halls from the viewpoint of acoustic potential of shapes. Acta Acustica United with Acustica, 102(6), 1027-1044

    2016

  10. [16]

    Lu, S., Xu, W., Chen, Y., & Yan, X. (2017). An experimental study on the acoustic absorp- tion of sand panels. Applied Acoustics, 116, 238-248

    2017

  11. [17]

    Supporting early -stage design decisions with building perfor- mance optimisation: Findings from a design experiment

    Lu S, Luo Y, Gao W, Lin B. Supporting early -stage design decisions with building perfor- mance optimisation: Findings from a design experiment. Journal of Building Engineering. 2024 Apr 1;82:108298

    2024

  12. [18]

    (2016, June)

    Lu, S., Yan, X., Xu, W., Chen, Y., & Liu, J. (2016, June). Improving auditorium designs with rapid feedback by integrating parametric models and acoustic simulation. In Building Simulation (Vol. 9, pp. 235-250). Tsinghua University Press

    2016

  13. [19]

    Street microclimate prediction based on Transformer model and street view image in high - density urban areas

    Ma, Xintong, Tiancheng Zeng, Miao Zhang, Pengyu Zeng, Borong Lin, and Shuai Lu. "Street microclimate prediction based on Transformer model and street view image in high - density urban areas." Building and Environment 269 (2025): 112490

    2025

  14. [20]

    Oommen, T. (2022). The glazed eyes of architectural history: Reflections on the (dis)contents of global history survey courses. Architectural Histories, 10(1), 1 –16. https://doi.org/10.16995/ah.8280

    2022 doi

  15. [21]

    Rousset, T., Kakibuchi , T., Sasaki, Y., & Nomura, Y. (2025). Merging Language and Do- main Specific Models: The Impact on Technical Vocabulary Acquisition. arXiv preprint arXiv:2502.12001

    2025 arXiv

  16. [23]

    (2025, April)

    Sun, H., Xia, B., Zhao, Y., Chang, Y., & Wang, X. (2025, April). Identical Human Prefer- ence Alignment Paradigm for Text -to-Image Models. In ICASSP 2025-2025 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 1-5). IEEE

    2025

  17. [24]

    (2025, April)

    Sun, H., Xia, B., Zhao, Y., Chang, Y., & Wang, X. (2025, April). Positive Enhanced Prefer- ence Alignment for Text -to-Image Models. In ICASSP 2025-2025 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 1-5). IEEE

    2025

  18. [25]

    Sun, H., Xia, B., Chang, Y., & Wang, X. (2025). Generalizing Alignment Paradigm of Text - to-Image Generation with Preferences Through f -Divergence Minimization. Proceedings of the AAAI Conference on Artificial Intelligence, 39(26), 27644 -27652. https://doi.org/10.1609/aaai....

    2025 doi

  19. [27]

    & Wang, X

    Wang, Y., He, Y., Wang, J., Li, K., Sun, L., Yin, J., ... & Wang, X. (2025). Enhancing intent understanding for ambiguous prompt: A human -machine co-adaption strategy. Neurocompu- ting, 130415. https://doi.org/10.1016/j.neucom.2024.130415

    2025

  20. [28]

    Wang, J., Liu, C., Zhu, W., He, Y., & Liu, Z. (2024). Beyond examples: High -level automat- ed reasoning paradigm in in -context learning via MCTS. arXiv preprint arXiv:2411.18478. https://arxiv.org/abs/2411.18478

    2024 arXiv

  21. [29]

    & Wang, X

    Wang, J., He, Y., Li, K., Li, S., Zhao, L., Yin, J., ... & Wang, X. (2025). MDANet: A multi - stage domain adaptation framework for generalizable low -light image enhancement. Neuro- computing, 129572

    2025

  22. [30]

    Yan, X., Lu, S., & Li, J. (2016). Experimental studies on the rain noise of lightweight roofs: Natural rains vs artificial rains. Applied Acoustics, 106, 63-76

    2016

  23. [31]

    (2024, May)

    Yin, J., Li, A., Xi, W., Yu, W., & Zou, D. (2024, May). Ground -fusion: A low -cost ground slam system robust to corner cases. In 2024 IEEE International Conference on Robotics and Automation (ICRA) (pp. 8603-8609). IEEE. 11

    2024

  24. [32]

    Yin, J., Gao, W., Li, J., Xu, P., Wu, C., Lin, B., & Lu, S. (2025). Archidiff: Interactive design of 3d architectural forms generated from a single image. Computers in Industry, 168, 104275

    2025

  25. [33]

    Yin, J., He, Y., Zhang, M., Zeng, P., Wang, T., Lu, S., & Wang, X. (2025). Promptlnet: Re- gion-adaptive aesthetic enhancement via prompt guidance in low -light enhancement net. arXiv preprint arXiv:2503.08276

    2025 arXiv

  26. [34]

    DRAG2BUILD: INTERACTIVE POINT - BASED MANIPULATION OF 3D ARCHITECTURAL POINT CLOUDS GENERATED FROM A SINGLE IM-AGE

    YIN, J., XU, P., GAO, W., ZENG, P., & LU, S. DRAG2BUILD: INTERACTIVE POINT - BASED MANIPULATION OF 3D ARCHITECTURAL POINT CLOUDS GENERATED FROM A SINGLE IM-AGE

  27. [35]

    Yu, T., Li, J., Jin, Y., Wu, W., Ma, X., Xu, W., & Lu, S. (2025). Machine learning prediction on spatial and environmental perception and work efficiency using electroencephalography including cross-subject scenarios. Journal of Building Engineering, 99, 111644

    2025

  28. [36]

    (2025, March)

    Zeng, T., Ma, X., Luo, Y., Yin, J., Ji, Y., & Lu, S. (2025, March). Improving outdoor thermal environmental quality through kinetic canopy empowered by machine learning and control algorithms. In Building Simulation (pp. 1-22). Beijing: Tsinghua University Press

    2025

  29. [37]

    Zeng, P., Gao, W., Li, J., Yin, J., Chen, J., & Lu, S. (2025). Automated residential layout generation and editing using natural language and images. Automation in Construction, 174, 106133

    2025

  30. [38]

    CARD: Cross -modal Agent Frame- work for Generative and Editable Residential Design

    Zeng, P., Jiang, M., Wang, Z., Li, J., Yin, J., & Lu, S. CARD: Cross -modal Agent Frame- work for Generative and Editable Residential Design. In NeurIPS 2024 Workshop on Open - World Agents

    2024

  31. [39]

    Zeng, P., Hu, G., Zhou, X., Li, S., & Liu, P. (2023). Seformer: a long sequence time -series forecasting model based on binary position encoding and information transfer regulariza- tion. Applied Intelligence, 53(12), 15747-15771

    2023

  32. [40]

    Zeng, P., Hu, G., Zhou, X., Li, S., Liu, P., & Liu, S. (2022). Muformer: A long sequence time-series forecasting model based on modified multi -head attention. Knowledge-Based Systems, 254, 109584

    2022

  33. [41]

    Zeng, P., Yin, J., Gao, Y., Li, J., Jin, Z., & Lu, S. (2025). Comprehensive and Dedicated Metrics for Evaluating AI -Generated Residential Floor Plans. Buildings, 15(10), 1674. https://doi.org/10.3390/buildings15101674

    2025 doi

  34. [42]

    Zhang, M., Yin, J., Zeng, P., Shen, Y., Lu, S., & Wang, X. (2025). TSCnet: A text -driven semantic-level controllable framework for customized low -light image enhancement. Neuro- computing, 129509

    2025

  35. [43]

    Zou, H., Ge, J., Liu, R., & He, L. (2023). Feature recognition of regional architecture forms based on machine learning: a case study of architecture heritage in Hubei Province, Chi- na. Sustainability, 15(4), 3504

    2023

  36. [44]

    Y., & Yin, J

    Zou, Y., Lou, S., Xia, D., Lun, I. Y., & Yin, J. (2021). Multi -objective building design opti- mization considering the effects of long -term climate change. Journal of Building Engineer- ing, 44, 102904

    2021

  37. [45]

    Zhang, Q., Li, Z., Chen, M., Liu, Y., Chen, Q., & Wang, Y. (2024). Pandora's Box or Alad- din's Lamp: A comprehensive analysis revealing the role of RAG noise in large language models. arXiv preprint arXiv:2408.13533. https://arxiv.org/abs/2408.13533

    2024 arXiv

  38. [46]

    Zhang, M., Shen, Y., Yin, J., Lu, S., & Wang, X. (2024). ADAGENT: Anomaly Detection Agent with Multimodal Large Models in Adverse Environments. IEEE Access

    2024

  39. [47]

    FloorPlan -DeepSeek (FPDS): A multimodal approach to floorplan genera tion using vector-based next room prediction

    Yin, Jun, et al. "FloorPlan -DeepSeek (FPDS): A multimodal approach to floorplan genera tion using vector-based next room prediction." arXiv preprint arXiv:2506.21562 (2025)

    2025 arXiv

  40. [48]

    UrbanSense: AFramework for Quantitative Analysis of Urban Streetscapes leveraging Vision Large Language Models

    Yin, Jun, et al. "UrbanSense: AFramework for Quantitative Analysis of Urban Streetscapes leveraging Vision Large Language Models." arXiv preprint arXiv:2506.10342 (2025)

    2025 arXiv

  41. [49]

    FloorplanMAE: A self-supervised framework for complete floorplan genera- tion from partial inputs

    Yin, Jun, et al. "FloorplanMAE: A self-supervised framework for complete floorplan genera- tion from partial inputs." arXiv preprint arXiv:2506.08363 (2025)

    2025 arXiv

  42. [50]

    Segment Any Architectural Facades (SAAF): An automatic segmentation model for building facades, walls and windows based on multimodal semantics guid- ance

    Li, Peilin, et al. "Segment Any Architectural Facades (SAAF): An automatic segmentation model for building facades, walls and windows based on multimodal semantics guid- ance." arXiv preprint arXiv:2506.09071 (2025)

    2025 arXiv

Pith tools

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