Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

SlideCoder: Layout-aware RAG-enhanced Hierarchical Slide Generation from Design

T0 review · 3 major / 5 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 SlideCoder generates editable slide code from reference images, beating state-of-the-art baselines by up to 40.5 points on the Slide2Code benchmark.

desk verdict Useful new benchmark and pipeline for reference-image-to-slide generation, but the training/benchmark overlap must be resolved before the headline numbers can be trusted. read the letter →

arxiv 2506.07964 v1 pith:6HWVXNKH submitted 2025-06-09 cs.CV cs.AI

Wenxin Tang , Jingyu Xiao , Wenxuan Jiang , Xi Xiao , Yuhang Wang , Xuxin Tang , Qing Li , Yuehe Ma
show 3 more authors
Junliang Liu Shisong Tang Michael R. Lyu
This is my paper
classification cs.CVcs.AI
keywords referenceimagetoslidegenerationSlide2Codebenchmarkretrieval-augmentedlayout-awarepromptcolorgradientsegmentationmultimodallargelanguagemodelspython-pptxcomplexitymetric
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 claims that converting a slide design image into an editable PowerPoint file can be made reliable by decomposing the image into blocks, describing each block in the vocabulary of the python-pptx object model, and retrieving exact API syntax at every step. To test this, the authors build Slide2Code, a difficulty-tiered benchmark of 300 slide images, and SlideCoder, a framework that combines color-gradient segmentation, a layout-aware prompt, and hierarchical retrieval-augmented generation. On Slide2Code, SlideCoder outperforms the strongest baseline by up to 40.5 points in an overall score, and the open-source 7B model SlideMaster approaches GPT-4o performance when used as the backbone. The authors argue that this makes vision-based, editable slide creation practical for layouts that natural-language prompts cannot describe.

What carries the argument

The load-bearing mechanism is the hierarchy of three components. CGSeg divides the reference image into a grid, marks blocks whose color-gradient magnitude exceeds a threshold, flood-fills connected regions, and recurses on each region to produce a tree of sub-images with coordinates. H-RAG supplies two knowledge bases retrieved at different stages: a Shape Type Knowledge Base maps visual elements to standardized python-pptx object descriptions, and an Operation Function Knowledge Base provides exact function signatures, parameters, and usage examples. The layout-aware prompt rescales block coordinates to absolute inches and feeds them, together with the global description and retrieved grammar, to the Assembler, which merges the generated snippets into a single executable script.

What would settle it

Release the Slide2Code sample identifiers and check them against the SLIDESBENCH training split used for SlideMaster; if any overlap exists, re-run the comparison on a strictly held-out set and see whether the 40.5-point overall margin and execution-rate gains persist.

Watch

Extended reading notes

Core claim

The central discovery is that reference-image-to-slide generation becomes tractable when a single monolithic generation step is replaced by a hierarchy: CGSeg recursively segments the design into visually coherent blocks; a Describer agent identifies each block's elements using python-pptx type names; a Coder agent emits per-block code snippets; and an Assembler agent merges snippets with absolute positioning and retrieved function grammars. The paper reports that this pipeline lifts execution success rates by up to 38 percentage points over the strongest baseline and improves both positional fidelity and visual similarity, and that the fine-tuned 7B SlideMaster model reaches an overall score of 78.8 on SLIDESBENCH, close to GPT-4o's 89.1 within the same SlideCoder framework.

Load-bearing premise

The 100 test slides per difficulty level are drawn from a pool that also fed SlideMaster's fine-tuning data, and the paper gives no split or exclusion to prove the test set and training set are disjoint, so a reader cannot yet rule out that some of the measured gain is memorization.

Editorial extensions

If this is right

  • Design-to-slide becomes a practical user workflow: a user can hand over a reference image and receive executable python-pptx code, not just a rendered mockup.
  • Execution success near 100% on simple-tier slides indicates that retrieval support can rescue small open models from library-syntax errors.
  • The Slide2Code difficulty tiers allow future systems to be compared separately on sparse, moderate, and dense layouts rather than on a single average.
  • Because SlideCoder wraps a generic MLLM backbone, any improvement in the underlying vision-language model should transfer directly to slide-generation quality.
  • The expanded reverse-engineering tool (10 object types, 44 styles) opens the door to building larger (image, code) training sets than AutoPresent's 5 types and 16 styles.

Reading between the lines

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

  • If the goal is a deployed assistant, the framework's assumption that the user provides separate design and picture components will need relaxing, since real reference images usually contain embedded raster content.
  • The benchmark draws its test pool from the same data universe used to fine-tune SlideMaster, so a documented split is needed before the reported margins can be read as generalization rather than memorization.
  • A natural extension is to evaluate whether the gains survive multi-slide generation with shared themes, since the current framework generates one slide at a time.
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

3 major / 5 minor

Summary. The paper formalizes the Reference Image (RI) to Slide Generation task, in which a slide design image is converted into editable Python-pptx code. It introduces the Slide2Code benchmark with difficulty tiers derived from a new Slide Complexity Metric (SCM), and proposes SlideCoder, a framework combining a Color Gradient-based Segmentation algorithm (CGSeg), a Layout-aware Prompt, and a Hierarchical Retrieval-Augmented Generation (H-RAG) scheme with three cooperative agents (Describer, Coder, Assembler). The paper also releases SlideMaster, a 7B model fine-tuned on reverse-engineered (RI, instruction, program) triplets. Experiments report large gains over AutoPresent baselines on Slide2Code and SLIDESBENCH, with execution rates, layout fidelity, visual metrics, and a human-correlation study for SCM. The code is released on GitHub.

Significance. If the claims hold, the paper would make a solid contribution: it defines a new vision-to-code task, provides a difficulty-tiered benchmark, proposes a modular retrieval-augmented pipeline, and releases an open-source 7B model approaching GPT-4o on this task. The evaluation is extensive, combining automated metrics, ablations, and a human correlation study, and the code release enables reproduction. The main caveat is that the central generalization claim depends on the Slide2Code evaluation samples being disjoint from SlideMaster's fine-tuning data, and the paper does not document this disjunction.

major comments (3)
  1. [§3.3 and §4.4] The paper does not document a split between the Slide2Code benchmark samples and the SLIDESBENCH data used to fine-tune SlideMaster. Section 3.3 constructs the sampling pool Y by drawing roughly 32,000 Zenodo10k instances and "additionally incorporat[ing] SLIDEBENCH samples in Y," then selects 100 samples per difficulty tier via KMeans clustering. Section 4.4 states that SlideMaster is fine-tuned on the "SLIDESBENCH training set." These are almost certainly the same corpus (the name appears as both SLIDEBENCH and SLIDESBENCH in the paper). If any of the 300 benchmark slides, or near-duplicates, appear in SlideMaster's training data, the headline gains (e.g., "approaching GPT-4o performance," up to 40.5-point improvements) partly measure memorization rather than generalization. The authors should either specify the exact exclusion procedure or re-run the evaluation on a provably disjoint subset.
  2. [§3.2, §4.1, §4.2.2] Several quantities that are load-bearing for reproducibility are not reported: the SCM weights α, β, γ (required in the weighted aggregation z_i = α·c̃_i + β·ẽ_i + γ·ṽ_i), the CGSeg parameters g, T, D_max (Algorithm 1), and the H-RAG retrieval top-k (Section 4.2.2). Without these values, the benchmark construction, the segmentation module, and the retrieval module cannot be reproduced or compared. Please report these hyperparameters in the main paper or appendices.
  3. [§5.4] The human validation of SCM is conducted on 100 samples randomly selected from the Slide2Code benchmark, but Slide2Code itself was constructed by KMeans-clustering SCM scores and sampling within clusters. This selection procedure can inflate the reported Pearson correlation (r = 0.873) because the benchmark deliberately spans the SCM score range. The authors should clarify whether the validation set is independent of the clustering/sampling step or, better, repeat the correlation analysis on a fresh random sample from the full pool Y. If the current design is retained, the conclusion that SCM is "a reliable and objective metric" should be tempered.
minor comments (5)
  1. [§3.2] There is a typo in the text: "To adress this" should be "To address this."
  2. [§3.3 and Figure 2] The dataset name is spelled inconsistently as both "SLIDEBENCH" and "SLIDESBENCH." Please use one spelling consistently and explicitly define which corpus is meant.
  3. [§1, Figure 1] The phrase "convert it into a editable slide" should be "convert it into an editable slide."
  4. [Table 1] The table layout is hard to parse because the color-coded difficulty levels are not reproduced in the text and the row grouping is unclear. Adding explicit level labels as column headers would improve readability.
  5. [§4.4] The LoRA hyperparameters are only given in Appendix C; please cite the table in the main text so readers can locate it.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity in the core SlideCoder/SlideMaster evaluation; the only self-referential element is the SCM human-validation study, which asks annotators to rate the same three dimensions the metric is built from.

  1. self definitional [Section 5.4 (Slide Complexity Metric Analysis); cf. Section 3.2 (Slide Complexity Metric) and Appendix D]
    "The annotators were instructed to score each slide from the perspective of three dimensions: the number of shapes, the diversity of shape types, and the level of element coverage. ... To assess the alignment between SCM and human perception, we first compute the Pearson correlation coefficient ... The result is r= 0.873 ..."

    SCM (Section 3.2) is defined as a weighted sum of exactly these three quantities: element count, element type count, and Element Coverage Ratio (vi, computed by CGSeg). The human-validity check therefore does not test SCM against an independent notion of complexity; it measures whether raters followed instructions to score the same three dimensions. A high correlation is substantially built into the protocol, so Section 5.4 cannot independently establish that SCM captures perceived complexity beyond its own definition. This is a secondary validation result, not the main SlideCoder-vs-baseline comparison, so it does not undermine the central empirical claims.

full rationale

The core claimed derivation is not circular. SlideCoder's components (CGSeg, H-RAG, layout-aware prompt, SlideMaster fine-tuning) are evaluated against Slide2Code using external metrics (CLIP, SSIM, content/position similarity, execution rate) that are not defined in terms of the framework's own outputs or fitted parameters. The reverse-engineered training data and the benchmark construction are separate pipelines; the benchmark provides reference images and PPTX files, while SlideMaster is trained to produce python-pptx code, and the evaluation compares rendered/structural outputs rather than checking whether the benchmark's own code was reproduced. No uniqueness theorem or load-bearing self-citation is invoked; citations to the authors' other UI-generation papers are related-work context, not premises. The one genuinely self-referential element is the SCM human-validation study: Section 5.4 asks annotators to rate the same three dimensions that Section 3.2 uses to define SCM, so the reported r=0.873 is partly a consistency check on the rating protocol rather than an external validation of the metric. Separately, the paper does not document whether the SLIDEBENCH samples incorporated into the Slide2Code pool (Section 3.3) are disjoint from the SLIDESBENCH training set used to fine-tune SlideMaster (Section 4.4); this is a potential train/test overlap and reproducibility risk, but it is not an equation-level circularity and cannot be proven from the text alone. Overall, the main comparative results remain independent content, so the circularity score is low.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a benchmark and a fine-tuned model. The main unpaid inputs are undisclosed SCM weights and CGSeg hyperparameters, the assumption that automated metrics measure quality, and the absence of a documented split between the Slide2Code evaluation pool and SlideMaster's SLIDESBENCH training source. No new physical or mathematical entities are introduced.

free parameters (5)
  • SCM aggregation weights (alpha, beta, gamma) = not reported (only alpha+beta+gamma=1)
    These weights determine the complexity score and difficulty tiers in Section 3.2 but are never given numerical values.
  • CGSeg grid size g = not reported
    Controls segmentation granularity in Algorithm 1 and affects block decomposition and prompts in Section 4.1.
  • CGSeg activation threshold T = not reported
    Blocks with gradient magnitude greater than T times the median are activated, so T controls segmentation behavior in Section 4.1.
  • CGSeg maximum recursion depth D_max = not reported
    Limits recursive decomposition and affects how fine-grained the resulting blocks are in Section 4.1.
  • H-RAG retrieval top-k = not reported
    The number of knowledge-base entries retrieved per prompt is unspecified in Section 4.2.2.
assumptions (4)
  • domain assumption Automated proxy metrics (CLIP, SSIM, content similarity, position similarity) are sufficient to rank slide-generation quality.
    Section 5.1 uses these as the evaluation without a separate user study of the generated slides.
  • domain assumption Slide2Code evaluation samples are disjoint from SlideMaster fine-tuning data.
    Assumed by the comparative evaluation in Section 5.2, but not explicitly guaranteed in Section 3.3 or Section 4.4.
  • domain assumption KMeans on normalized complexity dimensions yields well-separated simple, medium, and complex difficulty tiers.
    Section 3.3 relies on this for selecting benchmark samples without showing cluster validity.
  • domain assumption Reverse-engineered Python code from original PPTX files is a reliable training label.
    Sections 4.4 and 5.3 treat reconstruction ratio and CLIP score of the reverse tool as sufficient evidence of label quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SlideCoder: Layout-aware RAG-enhanced Hierarchical Slide Generation from Design." pith.science (2026). https://pith.science/paper/6HWVXNKH

@misc{pith2026250607964,
  author       = {Pith},
  title        = {Pith review of: SlideCoder: Layout-aware RAG-enhanced Hierarchical Slide Generation from Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6HWVXNKH}},
  note         = {Machine review of arXiv:2506.07964}
}
read the original abstract

Manual slide creation is labor-intensive and requires expert prior knowledge. Existing natural language-based LLM generation methods struggle to capture the visual and structural nuances of slide designs. To address this, we formalize the Reference Image to Slide Generation task and propose Slide2Code, the first benchmark with difficulty-tiered samples based on a novel Slide Complexity Metric. We introduce SlideCoder, a layout-aware, retrieval-augmented framework for generating editable slides from reference images. SlideCoder integrates a Color Gradient-based Segmentation algorithm and a Hierarchical Retrieval-Augmented Generation method to decompose complex tasks and enhance code generation. We also release SlideMaster, a 7B open-source model fine-tuned with improved reverse-engineered data. Experiments show that SlideCoder outperforms state-of-the-art baselines by up to 40.5 points, demonstrating strong performance across layout fidelity, execution accuracy, and visual consistency. Our code is available at https://github.com/vinsontang1/SlideCoder.

Figures

Figures reproduced from arXiv: 2506.07964 by the authors.

Figure 1
Figure 1. Illustration of slide generation scenarios from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Proportion of samples across three levels in the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The framework of SlideCoder. (a) Input Image (b) Activated Grid Blocks (c) Flood-filled Regions (d) Final result [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: An example of CGSeg applied to a slide reference image. The algorithm begins by computing color [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Examples of slides generated by different methods in three difficulty levels. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Evaluation guidelines provided to the four doctoral student annotators. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Prompt of Describer [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Prompt of Coder [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Layout-aware prompt [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Examples from the Shape Type knowledge base. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: An example from the Operation Function knowledge base. [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. EfficientUICoder: A Bidirectional Token Compression Framework for Efficient MLLM-Based UI Code Generation

    cs.SE 2025-09 conditional novelty 6.0 of 10

    EfficientUICoder compresses UI image tokens by 55-60% and suppresses duplicate HTML/CSS/text output, cutting LLM FLOPs by ~45% and inference time by ~49% on Llava-1.6-34b without quality loss on two benchmarks.

  2. When Cars Have Stereotypes: Auditing Demographic Bias in Objects from Text-to-Image Models

    cs.CV 2025-08 unverdicted novelty 6.0 of 10

    A new audit framework, SODA, measures demographic bias in objects generated by text-to-image models and finds strong default-to-majority and stereotype-collapse patterns across five models.

  3. Design First, Code Later: Aesthetically Pleasing Template-Free Slides Generation

    cs.HC 2026-04 conditional novelty 5.0 of 10

    A design-then-code multi-agent workflow, dataset, and small trained models produce template-free slides that humans prefer over strong baselines.

Reference graph

Works this paper leans on

40 extracted references · 14 canonical work pages · cited by 3 Pith papers

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    Shaikh Mostafa Al Masum, Mitsuru Ishizuka, and Md Tawhidul Islam. 2005. 'auto-presentation': a multi-agent system for building automatic multi-modal presentation of a topic from world wide web information. In IEEE/WIC/ACM International Conference on Intelligent Agent Technology, pages 246--249. IEEE

    work page 2005

  5. [5]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, and 1 others. 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923

    arXiv 2025

  6. [6]

    SV Burtsev and Ye P Kuzmin. 1993. An efficient flood-filling algorithm. Computers & graphics, 17(5):549--561

    work page 1993

  7. [7]

    Steve Canny. 2023. Python-ptx documentation

    work page 2023

  8. [8]

    Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. In Findings of the Association for Computational Linguistics ACL 2024, pages 2318--2335

    work page 2024

Show all 40 references
  1. [9]

    Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. 2009. Pearson correlation coefficient. Noise reduction in speech processing, pages 1--4

    2009

  2. [10]

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. 2024. The power of noise: Redefining retrieval for rag systems. In Proceedings of the 47th International ACM SIGIR Conference o...

    2024

  3. [11]

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,...

    2024

  4. [12]

    Tsu-Jui Fu, William Yang Wang, Daniel McDuff, and Yale Song. 2022. Doc2ppt: Automatic presentation slides generation from scientific documents. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 634--642

    2022

  5. [13]

    Jiaxin Ge, Zora Zhiruo Wang, Xuhui Zhou, Yi-Hao Peng, Sanjay Subramanian, Qinyue Tan, Maarten Sap, Alane Suhr, Daniel Fried, Graham Neubig, and 1 others. 2025. Autopresent: Designing structured visuals from scratch. arXiv preprint arXiv:2501.00912

    2025 arXiv

  6. [14]

    Yingqiang Ge, Wenyue Hua, Kai Mei, Juntao Tan, Shuyuan Xu, Zelong Li, Yongfeng Zhang, and 1 others. 2023. Openagi: When llm meets domain experts. Advances in Neural Information Processing Systems, 36:5539--5568

    2023

  7. [15]

    Google . 2025. Gemini API . https://ai.google.dev/gemini-api. Accessed: 2025-05-19

    2025

  8. [16]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  9. [17]

    Jun Han and Claudio Moraga. 1995. The influence of the sigmoid function parameters on the speed of backpropagation learning. In International workshop on artificial neural networks, pages 195--201. Springer

    1995

  10. [18]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718

    2021 arXiv

  11. [19]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

    2022

  12. [20]

    Yue Hu and Xiaojun Wan. 2014. Ppsgen: Learning-based presentation slides generation for academic papers. IEEE transactions on knowledge and data engineering, 27(4):1085--1097

    2014

  13. [21]

    Min-Yen Kan. 2007. Slideseer: A digital library of aligned document and presentation pairs. In Proceedings of the 7th ACM/IEEE-CS joint conference on Digital libraries, pages 81--90

    2007

  14. [22]

    Terry K Koo and Mae Y Li. 2016. A guideline of selecting and reporting intraclass correlation coefficients for reliability research. Journal of chiropractic medicine, 15(2):155--163

    2016

  15. [23]

    Kaixin Li, Yuchen Tian, Qisheng Hu, Ziyang Luo, Zhiyong Huang, and Jing Ma. 2024. Mmcode: Benchmarking multimodal large language models for code generation with visually rich programming problems. arXiv preprint arXiv:2404.09486

    2024 arXiv

  16. [24]

    Jenny GuangZhen Ma, Karthik Sreedhar, Vivian Liu, Pedro A Perez, Sitong Wang, Riya Sahni, and Lydia B Chilton. 2025. Dynex: Dynamic code synthesis with structured design exploration for accelerated exploratory programming. In Proceedings of the 2025 CHI Conference on Human Fac...

    2025

  17. [25]

    Daye Nam, Andrew Macvean, Vincent Hellendoorn, Bogdan Vasilescu, and Brad Myers. 2024. Using an llm to help with code understanding. In Proceedings of the IEEE/ACM 46th International Conference on Software Engineering, pages 1--13

    2024

  18. [26]

    Jim Nilsson and Tomas Akenine-M \"o ller. 2020. Understanding ssim. arXiv preprint arXiv:2006.13846

    2020 arXiv

  19. [27]

    Kunato Nishina and Yusuke Matsui. 2024. Svgeditbench: A benchmark dataset for quantitative assessment of llm's svg editing capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8142--8147

    2024

  20. [28]

    Juan A Rodriguez, Abhay Puri, Shubham Agarwal, Issam H Laradji, Sai Rajeswar, David Vazquez, Christopher Pal, and Marco Pedersoli. 2025. Starvector: Generating scalable vector graphics code from images and text. In Proceedings of the AAAI Conference on Artificial Intelligence,...

    2025

  21. [29]

    Athar Sefid and Jian Wu. 2019. Automatic slide generation for scientific papers. In Third International Workshop on Capturing Scientific Knowledge co-located with the 10th International Conference on Knowledge Capture (K-CAP 2019), SciKnow@ K-CAP 2019

    2019

  22. [30]

    Yuxuan Wan, Yi Dong, Jingyu Xiao, Yintong Huo, Wenxuan Wang, and Michael R Lyu. 2024. Mrweb: An exploration of generating multi-page resource-aware web code from ui designs. arXiv preprint arXiv:2412.15310

    2024 arXiv

  23. [31]

    Ronghuan Wu, Wanchao Su, and Jing Liao. 2024. Chat2svg: Vector graphics generation with large language models and image diffusion models. arXiv preprint arXiv:2411.16602

    2024

  24. [32]

    Jingyu Xiao, Yuxuan Wan, Yintong Huo, Zhiyao Xu, and Michael R Lyu. 2024. Interaction2code: How far are we from automatic interactive webpage generation? arXiv preprint arXiv:2411.03292

    2024

  25. [33]

    Jingyu Xiao, Ming Wang, Man Ho Lam, Yuxuan Wan, Junliang Liu, Yintong Huo, and Michael R. Lyu. 2025. https://arxiv.org/abs/2506.06251 Designbench: A comprehensive benchmark for mllm-based front-end code generation . Preprint, arXiv:2506.06251

    2025

  26. [34]

    Ximing Xing, Juncheng Hu, Guotao Liang, Jing Zhang, Dong Xu, and Qian Yu. 2024. Empowering llms to understand and generate complex vector graphics. arXiv preprint arXiv:2412.11102

    2024 arXiv

  27. [35]

    John Yang, Carlos E Jimenez, Alex L Zhang, Kilian Lieret, Joyce Yang, Xindi Wu, Ori Press, Niklas Muennighoff, Gabriel Synnaeve, Karthik R Narasimhan, and 1 others. 2024. Swe-bench multimodal: Do ai systems generalize to visual software domains? arXiv preprint arXiv:2410.03859

    2024 arXiv

  28. [36]

    Mingyue Yuan, Jieshan Chen, Zhenchang Xing, Aaron Quigley, Yuyu Luo, Tianqi Luo, Gelareh Mohammadi, Qinghua Lu, and Liming Zhu. 2024. Designrepair: Dual-stream design guideline-aware frontend repair with large language models. arXiv preprint arXiv:2411.01606

    2024 arXiv

  29. [37]

    Sukmin Yun, Haokun Lin, Rusiru Thushara, Mohammad Qazim Bhat, Yongxin Wang, Zutao Jiang, Mingkai Deng, Jinhong Wang, Tianhua Tao, Junbo Li, and 1 others. 2024. Web2code: A large-scale webpage-to-code dataset and evaluation framework for multimodal llms. arXiv preprint arXiv:2406.20098

    2024 arXiv

  30. [38]

    Fengji Zhang, Linquan Wu, Huiyu Bai, Guancheng Lin, Xiao Li, Xiao Yu, Yue Wang, Bei Chen, and Jacky Keung. 2024 a . Humaneval-v: Evaluating visual understanding and reasoning abilities of large multimodal models through coding tasks. arXiv preprint arXiv:2410.12381

    2024 arXiv

  31. [39]

    Linhao Zhang, Daoguang Zan, Quanshun Yang, Zhirong Huang, Dong Chen, Bo Shen, Tianyu Liu, Yongshun Gong, Pengjie Huang, Xudong Lu, and 1 others. 2024 b . Codev: Issue resolving with visual data. arXiv preprint arXiv:2412.17315

    2024 arXiv

  32. [40]

    Hao Zheng, Xinyan Guan, Hao Kong, Jia Zheng, Weixiang Zhou, Hongyu Lin, Yaojie Lu, Ben He, Xianpei Han, and Le Sun. 2025. Pptagent: Generating and evaluating presentations beyond text-to-slides. arXiv preprint arXiv:2501.03936

    2025 arXiv

Pith tools

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