Pith. sign in

REVIEW 4 major objections 4 minor 58 references

DrawAI: Agentic Benchmark and Workflow for Making Raster Images Editable

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper argues that turning raster images into editable files is chiefly a workflow problem: a two-stage agent pipeline raised Editability from 56.4 to 74.0 in all fourteen model-harness settings, while Fidelity moved only from 75.7 to 78

desk verdict Serious, well-built benchmark with a real workflow contribution, but the headline editability gain is inflated by a best-of-N comparison that isn't shown to be apples-to-apples. read the letter →

arxiv 2608.00548 v1 pith:C4DAUZ3F submitted 2026-08-01 cs.CV

Pu Cao , Qingye Kong , Xuedan Yin , Xuekun Zhao , Rupeng Yan , Qing Song , Yao Zhang , Lu Yang This is my paper
classification cs.CV
keywords image-to-editablereconstructionagenticworkflowraster-to-vectorSVGeditabilityfidelitymultimodalagentsbenchmark
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 introduces image-to-editable reconstruction: recovering a structured, directly manipulable file (SVG or PPTX) from a raster image such as a slide, poster, scientific figure, or diagram. It argues that visual Fidelity and structural Editability are distinct axes that usually trade off, and that most current systems maximize the first at the expense of the second. To study the task it contributes DrawAI-Bench, an 80-image benchmark scored by 39 hybrid criteria (deterministic metrics plus vision-language rubrics), and DrawAI-Flow, a two-stage agent workflow that first parses the image into an explicit reconstruction plan and then realizes the plan as executable graphics code through a code-render-validate-revise loop. The central empirical claim is that this structured workflow, not model choice, drives editable structure: across all fourteen model-harness settings Editability rose from 56.4 to 74.0 while Fidelity rose only from 75.7 to 78.7, and the workflow lifted the weakest configuration from 22.4 to 55.6 Overall. The improvement is not universal (67.5-90.0% of paired images improved per setting), but every setting gained on average, and a human study on 200 artifacts supports the benchmark's validity, with Overall scores correlating at rho = 0.94 with expert ratings of PPTX usability.

What carries the argument

Two mechanisms carry the argument. On the reconstruction side, DrawAI-Flow's Parser Agent fuses SAM3 region proposals and PP-OCRv5 text evidence into an explicit plan assigning each element an identity, geometry, appearance, and a reconstruction policy (native primitive vs. localized crop); the Reconstruction Agent then writes the artifact as an executable graphics program and iterates a code-render-validate-revise loop (up to five rounds), so revisions stay localized and the plan anchors the Fidelity-Editability trade-off to concrete choices. On the measurement side, the hybrid protocol — 39 criteria over six asset types (text, images, formulas, shapes, connectors, tables), 9 rule-based met

What would settle it

Re-score the 80 benchmark images' 30 rubric criteria with a different vision-language evaluator from another provider family (or an independent human panel), keeping the nine rule-based metrics fixed, then re-rank models. Because the top four configurations span only 1.4 Overall points while the workflow's Editability gain is about 17.6 points, material shifts in model ordering would expose evaluator bias, whereas the workflow-level claim would stand if the gain survives re-scoring.

Watch

Extended reading notes

Core claim

Central claim: image-to-editable reconstruction must be judged jointly on Fidelity and Editability, and an explicit two-stage workflow, not model choice, drives strong reconstructions. DrawAI-Flow's Parser Agent turns SAM3 region proposals and OCR text evidence into a per-element plan (native text/vector/formula object or localized crop); the Reconstruction Agent writes the artifact as executable graphics code and loops render-validate-revise up to five times. On 39 criteria, the workflow improved Overall in all fourteen model-harness settings, raised Editability by 17.6 points over direct prompting, and lifted the weakest configuration from 22.4 to 55.6. The paper also claims validity: expe

Load-bearing premise

Thirty of the benchmark's 39 criteria are scored by a single vision-language model, and the model-level rankings depend on that evaluator's agreement with human experts — measured on a 200-artifact sample at 80.1-93.3% exact agreement — extrapolating without bias across all 80 images and all model families.

Editorial extensions

If this is right

  • Editability is the prevailing deficit: ten of thirteen models score higher on Fidelity than Editability, and text is the weakest editable asset for nine of thirteen models, so any single-axis metric overstates authoring quality.
  • Workflow support substitutes for model capability at the weak end: the pipeline raised the weakest configuration from 22.4 to 55.6 Overall, gains are strongly anti-correlated with baseline strength (r = -0.89), and even the strongest direct agents gained 8.0-8.6 Overall points.
  • Harness choice participates in the Fidelity-Editability trade-off, not just in quality: one harness beat another for every model tested with both, and for one model it raised Fidelity by 8.2 points while lowering Editability by 5.3.
  • The workflow's gain is concentrated on structure: Text Editability rose +30.4 and Image Editability +26.9, while formula, shape, and connector Fidelity regressed slightly, indicating that plan-driven reconstruction converts recognized content into separable objects rather than uniformly solving geometry.
  • Cost and quality form a broad frontier rather than a single best point: nineteen settings span a 201x cost range over a 35.4-point quality range, and harness choice alone more than doubled cost at equal or lower quality for one strong model.

Reading between the lines

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

  • The workflow-level gain is so much larger than the model-level separations (top four configurations span 1.4 Overall points) that the central workflow claim would likely survive re-scoring with a different rubric evaluator, while the specific model rankings might not; the paper's own numbers make this asymmetry visible.
  • The asset taxonomy is format-agnostic enough that the same hybrid protocol could be extended to other editable targets such as HTML or native design-tool documents, which the paper does not test.
  • Because gains concentrate on text and image objects while formula/shape/connector fidelity regresses, the next bottleneck the authors leave implicit is geometric appearance — closing it may require direct visual-differentiation or differentiable rasterization rather than the current code loop.
  • Replacing the most reliable VLM rubrics with deterministic or rule-guided checks as evaluator models improve would make the benchmark cheaper and less dependent on a single vendor's model, a seam the hybrid design already leaves open.
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. DrawAI frames image-to-editable reconstruction: given a raster image, produce a structured artifact (SVG/PPTX) that preserves visual fidelity while exposing editable objects. DrawAI-Bench contains 80 images across scientific figures, slides, posters, and diagrams, with 39 evaluation criteria (9 rule-based, 30 rubric-based, the latter scored by GPT-5.6 Sol Ultra) aggregated into Fidelity, Editability, and Overall. DrawAI-Flow is a two-stage agentic workflow: a Parser Agent fuses SAM3 and OCR evidence into an explicit reconstruction plan, then a Reconstruction Agent synthesizes SVG via an iterative code-render-validate-revise loop. Experiments compare 13 models across 5 harnesses, a Simple Baseline, component ablations, specialized projects, and cost analysis. The central claims are that DrawAI-Bench tracks human judgment (PPTX-usability Spearman 0.88–0.94; VLM–human exact agreement 80.1–93.3%) and that DrawAI-Flow consistently improves editability (all 14 model–harness settings; mean Editability +17.6, Fidelity +3.0).

Significance. If the results hold, DrawAI makes a substantial methodological contribution to an emerging task: it gives the field a concrete task definition, an 80-image benchmark whose 39 criteria are fully operationalized in Appendix A with formulas and thresholds, and a genuine human-validation study with expert agreement bounds and PPTX-usability correlations. The paired per-image gain distributions in Tables 4 and 7 are a real strength, as is the transparent cost analysis. Three issues currently prevent full confidence in the headline claims: (1) the workflow-gain numbers in Table 6 depend on an asymmetric best-of-N collapse that is not documented for the Simple Baseline; (2) the component ablation suggests the workflow's benefit may be entirely attributable to off-the-shelf SAM3/OCR evidence; and (3) the model rankings rest on a single VLM evaluator from the same vendor family as the top-scoring models, with human validation not stratified by model family. These are fixable with additional analysis rather than fundamental flaws.

major comments (4)
  1. [§5.5, Table 6] The central workflow-gain claim rests on an asymmetric selection rule. The caption states 'Duplicate execution routes within a harness are collapsed to the higher-scoring DrawAI-Flow result,' but the paper reports neither the number of routes collapsed nor whether the Simple Baseline column received the same best-of-N treatment. Given the per-image ΔOverall SDs of 7–16 points in Tables 4 and 7, selecting the best of even two or three DrawAI-Flow runs can plausibly account for several points of the reported +17.6 Editability and +8–33 Overall gains. Please report the route counts, apply an identical selection rule to both arms (or use a single pre-registered run per condition), and provide paired confidence intervals for the headline workflow gains.
  2. [§5.6, Tables 9 vs 6] The component ablation undercuts the mechanism attributed to the workflow. With both OCR and SAM3 removed, GPT-5.5–Codex reaches 82.79 Overall (Table 9), essentially identical to the GPT-5.5–Codex Simple Baseline's 82.9 Overall (Table 6). Because the ablated condition still contains the Parser Agent, Reconstruction Agent, and the code–render–validate–revise loop, this indicates that the entire observed workflow gain is carried by the off-the-shelf OCR/SAM3 evidence rather than by 'explicit planning and materialization.' Please report the direct paired comparison between 'w/o OCR & SAM3' and the Simple Baseline on the same 80-image panel, and revise the mechanism discussion in Section 5.5/Conclusion accordingly.
  3. [§5.3, Tables 2 & 4] The headline model and harness rankings are presented without significance tests or confidence intervals. The top four Overall scores span only 1.4 points (GPT-5.5 about 90.95, Claude Opus 4.8 about 90.75, Kimi K3 about 90.2, GPT-5.6 Sol about 89.55), while per-image SDs in Tables 4 and 7 are 7–16 points. Similarly, §5.4 claims 'Claude Code outperforms OpenHands for every model' from mean matched-image gains whose SDs overlap zero substantially (e.g., MiMo V2.5 +1.4±17.1 in Table 4). Please provide bootstrap CIs or paired tests for the key model and harness comparisons; without them the rank ordering is not statistically established.
  4. [§3.2, §5.1–5.2] Thirty of 39 criteria are judged by a single VLM, GPT-5.6 Sol Ultra, whose vendor family also supplies the top Fidelity scorer (GPT-5.6 Sol) and the top Overall scorer (GPT-5.5). The human validation in Table 1 reports exact agreement of 80.1–93.3%, but only on a 200-artifact sample and not stratified by artifact provenance or by model family. The model-level rankings therefore presuppose that VLM–human agreement extrapolates to all 80 images and all 13 models without vendor-family bias. Please report VLM–human agreement separately for OpenAI-produced artifacts versus all others, and, if feasible, re-score the benchmark with a second evaluator from a different provider as a robustness check.
minor comments (4)
  1. [§5.2] The 200-artifact human-validation sample is described as spanning 'the full range of benchmark scores,' but the sampling protocol (e.g., stratified random by domain/provenance, quota targets) is not specified. Please clarify how the sample was selected and how many artifacts came from the DrawAI-Flow versus Simple Baseline conditions.
  2. [§5.3] The statement 'Model-level results select the strongest observed harness for each model' means Table 2 reports a best-of-harness upper envelope. This is a reasonable choice, but it can overstate a model's typical capability; consider also reporting per-model median or worst-harness results for transparency.
  3. [§5.7] The text says 'The former exhaustive 39-row project comparison table is omitted' and then claims DrawAI obtains the best value on 35 of 39 criteria. This claim is not verifiable if the table is omitted. Please include the 39-row comparison in the supplementary material.
  4. [Appendix A, Table 12] Some rule-based criteria depend on unspecified preprocessing: 'Split SVG text lines may be grouped before text-box IoU is computed' and 'unmatched ground-truth image boxes contribute zero.' Please specify the grouping algorithm and matching procedure precisely so the metrics are fully reproducible.

Circularity Check

1 steps flagged · score 4.0 of 10

Workflow-gain comparison is partly selected on the evaluation scores it then reports; the rest of the derivation is self-contained.

  1. fitted input called prediction [Table 6 caption; Section 5.5]
    "Duplicate execution routes within a harness are collapsed to the higher-scoring DrawAI-Flow result. ... Across all paired outputs, Fidelity rises from 75.7 to 78.7, whereas Editability rises from 56.4 to 74.0"

    The DrawAI-Flow score used in the paired workflow comparison is the maximum over duplicate routes, while the Simple Baseline column is not reported as being selected the same way. Thus the headline Editability gain (+17.6) and the claim that all fourteen settings improve are partly forced by the selection rule: taking the maximum of several attempts raises the score even if no single DrawAI-Flow run is better than the baseline. The evaluation scores are used both to choose which route to report and as the measure of the workflow's effect, so the reported 'prediction' that DrawAI-Flow improves editable structure is statistically inflated by construction. The paper does not state the number of collapsed routes or apply the same best-of-N rule to the baseline arm, so the gain is not a clean s

full rationale

The paper's benchmark itself is not circular: DrawAI-Bench is human-validated on 200 artifacts, with VLM-human exact agreement of 80.1-93.3% and Spearman correlations of 0.88-0.94 against PPTX usability, and the rule-based metrics are deterministic. There are no load-bearing self-citations or imported uniqueness theorems; the self-citations in Related Work are background references. However, the central claim that DrawAI-Flow 'consistently improves editable structure' rests on Table 6, whose caption discloses that duplicate DrawAI-Flow routes are collapsed to the higher-scoring result without stating that the Simple Baseline received the same treatment. This is a partial reduction: the workflow's gain is at least partly an artifact of selecting the best DrawAI-Flow run using the same evaluation scores that are then reported as the workflow's quality. Because the workflow's benefit may survive under a fair single-run comparison (the ablations in Section 5.6 show systematic effects), the circularity is confined to the headline gain magnitude rather than the entire derivation, warranting a score of 4 rather than higher.

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

This is an empirical benchmark paper with no derivation, so the ledger records the hand-chosen evaluation thresholds, the equal-weight aggregation, the VLM answer scale, and the iteration cap, together with the annotation-quality, sample-representativeness, and evaluator-validity assumptions on which the measurements rest. All are disclosed in the paper (Appendix A, Sections 3.1, 3.2, 4.3, 5.2). No new physical or conceptual entities are posited: the Parser Agent, Reconstruction Agent, element registry, and reconstruction plan are software components, not postulated entities carrying an independent-evidence burden.

free parameters (4)
  • Rule-based thresholds (text similarity 0.8, text-box IoU 0.5, image IoU 0.5, region MSE 0.02) = 0.8 / 0.5 / 0.5 / 0.02
    Appendix A (page 21): these hand-chosen defaults define matching for text content, text-box IoU, image mIoU, and the localized-image fallback, and directly set the deterministic criterion scores.
  • Overall weighting and aggregation (equal Fidelity/Editability weights, hierarchical macro-averaging) = 0.5 / 0.5; macro-average over categories
    Section 3.2: 'compute Overall as the mean of Fidelity and Editability'. Different weights would reorder model rankings, so the equal weighting is a free choice the headline results depend on.
  • VLM rubric answer mapping (yes=1, partial=0.5, no=0) = 1 / 0.5 / 0
    Section 3.2 and Eq. 2 in Appendix A: this discrete scale affects all 30 rubric criteria and therefore both headline dimensions.
  • DrawAI-Flow iteration cap = 5 rounds
    Section 4.3: 'in each of the maximum five rounds'. More rounds would likely change both quality and cost, so the cap is a design choice embedded in every reported result.
assumptions (5)
  • domain assumption Ground-truth element annotations G_i (GPT-5.5 initial, manually corrected) are accurate to the level required by bbox-IoU and rubric criteria.
    Section 3.1: annotations determine which criteria apply and provide element-level references; systematic annotation errors would bias every system's scores.
  • domain assumption The 80-image corpus (20 per domain, 10 real and 10 AI-generated) is representative of practical visual-creation scenarios.
    Section 3.1 states this as design intent; with n=80 and per-artifact SDs of 7-19 points (Table 11), the generalizability of the headline rankings rests on this assumption.
  • domain assumption GPT-5.6 Sol Ultra rubric judgments approximate expert human judgment on all evaluated artifacts, not only the 200-artifact validation sample.
    Section 5.2 validates on a balanced sample (exact agreement 80.1-93.3%); the full benchmark numbers in Tables 2 and 3 assume this agreement extrapolates to all images and models.
  • domain assumption The deterministic renderer RENDER(A_i) is a faithful proxy for how a user sees and edits the artifact.
    Section 3.2: all comparisons use rendered SVGs; renderer-specific artifacts would propagate into every criterion score.
  • domain assumption Expert majority vote is the ground truth for perceptual and semantic rubric criteria.
    Section 5.2: consensus is formed by majority vote after a shared calibration round; inter-expert alpha of 0.66-0.78 shows only moderate agreement, so majority vote is a noisy ground truth.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DrawAI: Agentic Benchmark and Workflow for Making Raster Images Editable." pith.science (2026). https://pith.science/paper/C4DAUZ3F

@misc{pith2026260800548,
  author       = {Pith},
  title        = {Pith review of: DrawAI: Agentic Benchmark and Workflow for Making Raster Images Editable},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4DAUZ3F}},
  note         = {Machine review of arXiv:2608.00548}
}
read the original abstract

Recent image-generation models and multimodal agents can produce high-quality visuals for increasingly complex visual communication tasks. Yet their raster outputs remain difficult to use directly because meaningful content and relationships are flattened into pixels, preventing users from inspecting, modifying, rearranging, or reusing individual components. We formulate image-to-editable reconstruction, which recovers a structured, directly manipulable artifact from a raster image while preserving its visual and semantic content. The central challenge is to jointly satisfy Fidelity and Editability, which often trade off in practice. To study this task, we introduce DrawAI, comprising an agentic benchmark, DrawAI-Bench, and a reconstruction workflow, DrawAI-Flow. DrawAI-Bench spans scientific figures, presentation slides, posters, and diagrams, combining real and AI-generated images to reflect practical visual-creation scenarios. It evaluates Fidelity and Editability through a hybrid protocol of 39 criteria: deterministic rule-based metrics measure properties with direct correspondences, while asset-specific vision-language rubrics capture semantic and perceptual qualities for which exact matching is misleading. Besides, we propose DrawAI-Flow, a two-stage agentic workflow in which a Parser Agent turns extracted elements evidence into an explicit reconstruction plan, and a Reconstruction Agent realizes the plan as executable graphics code through an iterative code-render-validate-revise loop. On DrawAI-Bench, we systematically evaluate thirteen models across five agent harnesses to study the effects of model capability, harness choice, and workflow design. The results show that reconstruction quality and costs vary substantially across model-harness configurations, while DrawAI-Flow consistently improves editable structure.

Figures

Figures reproduced from arXiv: 2608.00548 by the authors.

Figure 8
Figure 8. Markdown prompt template used by the Simple Baseline for image-to-editable-SVG reconstruction. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 27 canonical work pages

  1. [1]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. InProceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of Machine Learning Research, pages 8821–8831. PMLR, 2021. URL https: //proceedings.mlr.press/v139/ramesh21a.html

    work page 2021

  2. [2]

    GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models

    Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. InProceedings of the 39th International Conference on Machine Learning, volume 162 ofProceedings of Machine Learning Research, pages...

    work page 2022

  3. [3]

    Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Moham- mad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. InAd- vances in Neural Information Processing Systems, volume 3...

    work page

  4. [5]

    Paperbanana: Automat- ing academic illustration for ai scientists, 2026

    Dawei Zhu, Rui Meng, Yale Song, Xiyu Wei, Sujian Li, Tomas Pfister, and Jinsung Yoon. Paperbanana: Automat- ing academic illustration for ai scientists, 2026. URL https://arxiv.org/abs/2601.23265

    arXiv 2026

  5. [6]

    Autofigure: Generating and refining publication-ready scientific illustrations

    Minjun Zhu, Zhen Lin, Yixuan Weng, Panzhong Lu, Qiujie Xie, Yifan Wei, Sifan Liu, Qiyao Sun, and Yue Zhang. Autofigure: Generating and refining publication-ready scientific illustrations. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=5N3z9JQJKq

    work page 2026

  6. [7]

    ReAct: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations,

  7. [8]

    MM-REACT: Prompting ChatGPT for multimodal reasoning and action, 2023

    Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. MM-REACT: Prompting ChatGPT for multimodal reasoning and action, 2023. URL https://arxiv.org/abs/2303.11381

    arXiv 2023

  8. [9]

    Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, et al

    Xingyao Wang, Boxuan Li, Yufan Song, Frank F. Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, et al. Openhands: An open platform for ai software developers as generalist agents, 2024. URL https://arxiv.org/abs/2407.16741

    arXiv 2024

Show all 58 references
  1. [10]

    Executable code actions elicit better LLM agents

    Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. Executable code actions elicit better LLM agents. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, pages 50208–50...

    2024

  2. [12]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthes...

    2024

  3. [13]

    Qwen-image technical report, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai...

    2025 arXiv

  4. [14]

    Gligen: Open-set grounded text-to-image generation

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 22511–22521, 2023. doi:...

    2023

  5. [15]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836–3847, 2023. URL https://openaccess.thecvf.com/content/ICCV2023/html/Zhang_Adding_Con...

    2023

  6. [16]

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models.Proceedings of the AAAI Conference on Artificial Intelligence, 38(5):4296–4304, 2024. ...

    2024 doi

  7. [17]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models, 2023

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models, 2023. URL https://arxiv.org/abs/2308.06721

    2023 arXiv

  8. [18]

    Controllable generation with text-to-image diffusion models: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 48(4):4771–4791, 2026

    Pu Cao, Feng Zhou, Qing Song, and Lu Yang. Controllable generation with text-to-image diffusion models: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 48(4):4771–4791, 2026. doi: 10.1109/TPAMI. 2025.3646548. URL https://doi.org/10.1109/TPAMI.2025.3646548

    2026

  9. [19]

    Image is all you need to empower large-scale diffusion models for in-domain generation

    Pu Cao, Feng Zhou, Lu Yang, Tianrui Huang, and Qing Song. Image is all you need to empower large-scale diffusion models for in-domain generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18358–18368, 2025. doi: 10.1109/CVPR5273...

    2025

  10. [20]

    Rodriguez, David Vazquez, Issam Laradji, Marco Pedersoli, and Pau Rodriguez

    Juan A. Rodriguez, David Vazquez, Issam Laradji, Marco Pedersoli, and Pau Rodriguez. Figgen: Text to scientific figure generation, 2023. URL https://arxiv.org/abs/2306.00800

    2023 arXiv

  11. [21]

    Autofigure-edit: Generating editable scientific illustrations via reference-guided styling

    Zhen Lin, Qiujie Xie, Minjun Zhu, Shichen Li, QiYao Sun, Enhao Gu, Yiran Ding, Ke Sun, Fang Guo, Panzhong Lu, Zhiyuan Ning, Yixuan Weng, and Yue Zhang. Autofigure-edit: Generating editable scientific illustrations via reference-guided styling. InProceedings of the 64th Annual ...

    2026 doi

  12. [22]

    Crafter: A multi-agent harness for editable scientific figure generation from diverse inputs, 2026

    Haozhe Zhao, Shuzheng Si, Zhenhailong Wang, Zheng Wang, Liang Chen, Xiaotong Li, Zhixiang Liang, Maosong Sun, and Minjia Zhang. Crafter: A multi-agent harness for editable scientific figure generation from diverse inputs, 2026. URL https://arxiv.org/abs/2605.30611

    2026 arXiv

  13. [23]

    From pixels to paths: A multi-agent framework for editable scientific illustration, 2025

    Jianwen Sun, Fanrui Zhang, Yukang Feng, Chuanhao Li, Zizhen Li, Jiaxin Ai, Yifan Chang, Yu Dai, and Kaipeng Zhang. From pixels to paths: A multi-agent framework for editable scientific illustration, 2025. URL https://arxiv.org/abs/2510.27452

    2025

  14. [24]

    Doc2ppt: Automatic presentation slides generation from scientific documents.Proceedings of the AAAI Conference on Artificial Intelligence, 36(1):634–642,

    Tsu-Jui Fu, William Yang Wang, Daniel McDuff, and Yale Song. Doc2ppt: Automatic presentation slides generation from scientific documents.Proceedings of the AAAI Conference on Artificial Intelligence, 36(1):634–642,

  15. [25]

    Pptagent: Generating and evaluating presentations beyond text-to-slides

    Hao Zheng, Xinyan Guan, Hao Kong, Wenkai Zhang, Jia Zheng, Weixiang Zhou, Hongyu Lin, Yaojie Lu, Xianpei Han, and Le Sun. Pptagent: Generating and evaluating presentations beyond text-to-slides. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Proc...

    2025 doi

  16. [26]

    Paper2poster: Towards multimodal poster automation from scientific papers

    Wei Pang, Kevin Qinghong Lin, Xiangru Jian, Xi He, and Philip Torr. Paper2poster: Towards multimodal poster automation from scientific papers. InAdvances in Neural Information Processing Systems, 2025. URL https://arxiv.org/abs/2505.21497

    2025

  17. [27]

    URL https://ojs.aaai.org/index.php/AAAI/article/view/19943

    doi: 10.1609/aaai.v36i1.19943. URL https://ojs.aaai.org/index.php/AAAI/article/view/19943. 17

    doi

  18. [28]

    Presentbench: A fine-grained rubric-based benchmark for slide generation, 2026

    Xin-Sheng Chen, Jiayu Zhu, Pei-lin Li, Hanzheng Wang, Shuojin Yang, and Meng-Hao Guo. Presentbench: A fine-grained rubric-based benchmark for slide generation, 2026. URL https://arxiv.org/abs/2603.07244

    2026

  19. [29]

    Deckbench: Benchmarking multi-agent frameworks for academic slide generation and editing, 2026

    Daesik Jang, Morgan Lindsay Heisler, Linzi Xing, Yifei Li, Edward Wang, Ying Xiong, Yong Zhang, and Zhenan Fan. Deckbench: Benchmarking multi-agent frameworks for academic slide generation and editing, 2026. URL https://arxiv.org/abs/2602.13318

    2026 arXiv

  20. [30]

    Slidesgen-bench: Evaluating slides generation via computational and quantitative metrics, 2026

    Yunqiao Yang, Wenbo Li, Houxing Ren, Zimu Lu, Ke Wang, Zhiyuan Huang, Zhuofan Zong, Mingjie Zhan, and Hongsheng Li. Slidesgen-bench: Evaluating slides generation via computational and quantitative metrics, 2026. URL https://arxiv.org/abs/2601.09487

    2026

  21. [31]

    Chartmimic: Evaluating LMM’s cross-modal reasoning capability via chart-to-code generation

    Cheng Yang, Chufan Shi, Yaxin Liu, Bo Shui, Junjie Wang, Mohan Jing, Linran Xu, Xinyu Zhu, Siheng Li, Yuxiang Zhang, Gongye Liu, Xiaomei Nie, Deng Cai, and Yujiu Yang. Chartmimic: Evaluating LMM’s cross-modal reasoning capability via chart-to-code generation. InInternational C...

    2025 arXiv

  22. [32]

    pix2code: Generating code from a graphical user interface screenshot, 2017

    Tony Beltramelli. pix2code: Generating code from a graphical user interface screenshot, 2017. URL https: //arxiv.org/abs/1705.07962

    2017 arXiv

  23. [33]

    Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M. Rush. Image-to-markup generation with coarse- to-fine attention. InProceedings of the 34th International Conference on Machine Learning, volume 70 ofProceedings of Machine Learning Research, pages 980–989. PMLR, 20...

    2017

  24. [34]

    Xing, Xiaodan Liang, and Zhiqiang Shen

    Sukmin Yun, Haokun Lin, Rusiru Thushara, Mohammad Qazim Bhat, Yongxin Wang, Zutao Jiang, Mingkai Deng, Jinhong Wang, Tianhua Tao, Junbo Li, Haonan Li, Preslav Nakov, Timothy Baldwin, Zhengzhong Liu, Eric P . Xing, Xiaodan Liang, and Zhiqiang Shen. Web2code: A large-scale webpa...

    2024 doi

  25. [35]

    Differentiable vector graphics rasteri- zation for editing and learning.ACM Transactions on Graphics, 39(6):1–15, 2020

    Tzu-Mao Li, Michal Lukᡠc, Michaël Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasteri- zation for editing and learning.ACM Transactions on Graphics, 39(6):1–15, 2020. doi: 10.1145/3414685.3417871. URL https://doi.org/10.1145/3414685.3417871

    2020

  26. [36]

    Design2code: Benchmarking multimodal code generation for automated front-end engineering

    Chenglei Si, Yanzhe Zhang, Ryan Li, Zhengyuan Yang, Ruibo Liu, and Diyi Yang. Design2code: Benchmarking multimodal code generation for automated front-end engineering. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computatio...

    2025 doi

  27. [37]

    Towards layer-wise image vectorization

    Xu Ma, Yuqian Zhou, Xingqian Xu, Bin Sun, Valerii Filev, Nikita Orlov, Yun Fu, and Humphrey Shi. Towards layer-wise image vectorization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16314–16323, 2022. doi: 10.1109/CVPR52688.2022.01...

    2022

  28. [38]

    Rosin, and Yu-Kun Lai

    Teng Hu, Ran Yi, Baihong Qian, Jiangning Zhang, Paul L. Rosin, and Yu-Kun Lai. Supersvg: Superpixel-based scalable vector graphics synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24892–24901, 2024. doi: 10.1109/CVPR52733.20...

    2024

  29. [39]

    Omnisvg: A unified scalable vector graphics generation model

    Yiying Yang, Wei Cheng, Sijin Chen, Xianfang Zeng, Fukun Yin, Jiaxu Zhang, Liao Wang, Gang Yu, Xingjun Ma, and Yu-Gang Jiang. Omnisvg: A unified scalable vector graphics generation model. InAdvances in Neural Information Processing Systems, 2025. URL https://arxiv.org/abs/2504.06263

    2025

  30. [40]

    Deepsvg: A hierarchical generative network for vector graphics animation

    Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. Deepsvg: A hierarchical generative network for vector graphics animation. InAdvances in Neural Information Processing Systems, volume 33, 2020. URL https://proceedings.neurips.cc/paper/2020/hash/bcf9d6bd14...

    2020

  31. [41]

    Semantic document derendering: SVG reconstruction via vision-language modeling.Proceedings of the AAAI Conference on Artificial Intelligence, 40(6):4636–4644, 2026

    Adam Hazimeh, Ke Wang, Mark Collier, Gilles Baechler, Efi Kokiopoulou, and Pascal Frossard. Semantic document derendering: SVG reconstruction via vision-language modeling.Proceedings of the AAAI Conference on Artificial Intelligence, 40(6):4636–4644, 2026. doi: 10.1609/aaai.v4...

    2026 doi

  32. [42]

    From dead pixels to editable slides: Infographic reconstruction into native Google Slides via vision-language region understanding

    Leonardo Gonzalez. From dead pixels to editable slides: Infographic reconstruction into native Google Slides via vision-language region understanding. InCompanion Proceedings of the ACM Web Conference 2026, 2026. URL https://arxiv.org/abs/2602.07645

    2026

  33. [43]

    Amodalsvg: Amodal image vectorization via semantic layer peeling, 2026

    Juncheng Hu, Ziteng Xue, Guotao Liang, Anran Qi, Buyu Li, Sheng Wang, Dong Xu, and Qian Yu. Amodalsvg: Amodal image vectorization via semantic layer peeling, 2026. URL https://arxiv.org/abs/2604.10940

    2026 arXiv

  34. [44]

    Svgeditbench: A benchmark dataset for quantitative assessment of LLM’s SVG editing capabilities

    Kunato Nishina and Yusuke Matsui. Svgeditbench: A benchmark dataset for quantitative assessment of LLM’s SVG editing capabilities. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2024. URL https://arxiv.org/abs/2404.13710

    2024 arXiv

  35. [45]

    Vcg-bench: Towards a unified visual-centric benchmark for structured generation and editing, 2026

    Xiaoyan Su, Peijie Dong, Zhenheng Tang, Song Tang, Yuyao Zhai, Kaitao Lin, Liang Chen, Yuhang Gai, Yuyu Luo, Qiang Wang, and Xiaowen Chu. Vcg-bench: Towards a unified visual-centric benchmark for structured generation and editing, 2026. URL https://arxiv.org/abs/2605.15677

    2026 arXiv

  36. [46]

    Graphic-design-bench: A comprehensive benchmark for evaluating AI on graphic design tasks, 2026

    Adrienne Deganutti, Elad Hirsch, Haonan Zhu, Jaejung Seol, and Purvanshi Mehta. Graphic-design-bench: A comprehensive benchmark for evaluating AI on graphic design tasks, 2026. URL https://arxiv.org/abs/2604. 04192

    2026

  37. [47]

    Canvas: A benchmark for vision- language models on tool-based user interface design.Proceedings of the AAAI Conference on Artificial Intelligence, 40(26):22182–22190, 2026

    Daeheon Jeong, Seoyeon Byun, Kihoon Son, Dae Hyun Kim, and Juho Kim. Canvas: A benchmark for vision- language models on tool-based user interface design.Proceedings of the AAAI Conference on Artificial Intelligence, 40(26):22182–22190, 2026. doi: 10.1609/aaai.v40i26.39374. URL...

    2026 doi

  38. [48]

    Anchorflow: Editable SVG reconstruction via sparse anchor point fields, 2026

    Mengnan Jiang, Christian Franke, Michele Franco Adesso, Antonio Haas, and Grace Li Zhang. Anchorflow: Editable SVG reconstruction via sparse anchor point fields, 2026. URL https://arxiv.org/abs/2605.19551

    2026 arXiv

  39. [49]

    PaddleOCR 3.0 technical report, 2025

    Cheng Cui, Ting Sun, Manhui Lin, Tingquan Gao, Yubo Zhang, Jiaxuan Liu, Xueqing Wang, Zelun Zhang, Changda Zhou, Hongen Liu, et al. PaddleOCR 3.0 technical report, 2025. URL https://arxiv.org/abs/2507.05595. 19 DrawAI: Supplementary Material A Complete Evaluation Criteria This...

    2025 arXiv

  40. [52]

    SAM 3: Segment anything with concepts,

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. SAM 3: Segment anything with concepts,

  41. [55]

    Inspect the complete source image and infer its canvas dimensions, visual hierarchy, major regions, text, spatial relationships, colors, and reusable graphical structures

  42. [56]

    Distinguish content that should be represented as editable SVG text and vector primitives from complex raster content that should be preserved through localized source crops

    Decompose the slide into semantic components. Distinguish content that should be represented as editable SVG text and vector primitives from complex raster content that should be preserved through localized source crops

  43. [57]

    Set the SVG viewBox to match the source dimensions or an exactly proportional coordinate system, then reconstruct the slide in visual stacking order: background, large regions, structural elements, text, diagrams, and localized raster assets

  44. [58]

    Preserve grouping, alignment, line breaks, emphasis, and approximate font metrics

    Preserve text as SVG <text> and <tspan> elements when it is readable. Preserve grouping, alignment, line breaks, emphasis, and approximate font metrics

  45. [59]

    When local deterministic SVG rendering is available, render output/result.svg to output/preview.png, compare it with the source image, and revise discrepancies in element presence, layout, text placement, scale, color, style, stacking order, and clarity

  46. [60]

    Generate the final SVG only after checking that it is valid, standalone, renderable, and visually faithful. Fidelity Criteria • Preserve the source slide's aspect ratio, composition, alignment, spacing, typography hierarchy, colors, borders, and relative element sizes as close...

  47. [2022]

    URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/ ec795aeadae0b7d230fa35cbaf04c041-Abstract-Conference.html

    doi: 10.52202/068431-2643. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/ ec795aeadae0b7d230fa35cbaf04c041-Abstract-Conference.html

    2022 doi

  48. [2023]

    URL https://arxiv.org/abs/2210.03629

    arXiv

  49. [2025]

    URL https://arxiv.org/abs/2511.16719

    arXiv

  50. [3560]

    URL https://proceedings.neurips.cc/paper_files/paper/2024/hash/cb66be286795d71f89367d596bf78ea7- Abstract-Datasets_and_Benchmarks_Track.html

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.