Pixels or Code? Benchmarking Instruction-Guided Visual Artifact Operation

Shun Guo; Shuizhou Shi; Boao Wang; Wei Xiong; Siyu Sun; Jiangtong Li; Jie Li

cs.CVOct 12, 2026

Pixels or Code? Benchmarking Instruction-Guided Visual Artifact Operation

Authors: Shun Guo, Shuizhou Shi, Boao Wang, Wei Xiong, Siyu Sun, Jiangtong Li, Jie Li

Organizations: Tongji University · Shanghai Jiao Tong University

Abstract

Visual artifacts such as charts, technical diagrams, webpages, and document pages underpin scientific, commercial, and engineering workflows. Editing these artifacts requires direct modification, dependency-aware updates, preservation of unaffected content, and compliance with domain constraints. We define this capability as Visual Artifact Operation (VAO). Prior benchmarks test parts of this problem, including cascading updates in charts. We study direct edits, dependent updates, and preservation across visual domains and pixel, code, and agent paradigms under one image-only protocol. We introduce Visual Artifact Operation Bench (VAO-Bench), comprising 1,573 manually screened tasks across nine artifact categories and three difficulty levels. VAO-Bench uses executable artifacts and deterministic rendering to trace reference transformations and annotate direct edits, dependent updates, and preservation constraints. All systems receive the same source image and instruction; we evaluate pixel-space editors, code-generating multimodal systems, and tool-using agents on direct-edit accuracy, dependency-aware update, content preservation, domain validity, and visual quality. Experiments show that executable-code systems better support dependent updates, while pixel-space editors offer faster execution and strong visual quality. Explicitly stating dependencies improves both paradigms, identifying implicit-update inference as a central bottleneck. VAO-Bench provides a reproducible foundation for dependable visual artifact editing across generation paradigms.

Figures & tables

Figure 1: Visual artifact operation (left) and benchmark comparison (right). Size follows each source’s unit. Code counts output code environments, not source-code access. Sci., scientific; UI, interfaces; Diag., diagrams; Diff., difficulty levels or task-complexity analysis; Deps., dependent-update evaluation ( $\checkmark$ : explicit; $\odot$ : related checks; $\times$ : absent). Best viewed by zoom in.

Figure 2: Three-stage VAO-Bench construction pipeline: the language model generates specifications, instructions, annotations, and plans; the multimodal model generates and edits visual code and reviews its renderings. Failed checks allow up to three repair rounds. Accepted packages retain code, images, metadata, annotations, and verification records.

Figure 3: Dataset composition and task complexity of VAO-Bench. The panels summarize artifact-category and rendering-family counts, difficulty by category, and per-task score points for direct changes, dependent updates, and content preservation.

Track	Method	Time (s)	Comp.	Auto-zero					Completed-only
Track	Method	Time (s)	Comp.	Dir.	Der.	Inv.	VQ	DQ	Dir.	Der.	Inv.	VQ	DQ
Direct	GPT-Image-2	154.2	100%	0.730	0.527	0.916	4.47	3.79	0.730	0.527	0.916	4.47	3.79
Direct	Seedream-5.0-Pro	226.8	94.0%	0.756	0.622	0.890	4.20	3.81	0.804	0.661	0.949	4.47	4.05
Native	GPT-5.6-Sol	300	97.2%	0.806	0.664	0.925	4.06	4.10	0.831	0.683	0.951	4.17	4.22
	DeepSeek-Flash	354	94.7%	0.778	0.631	0.826	3.99	3.81	0.825	0.671	0.874	4.21	4.02
	GLM-5.3-Flash	1353	50.4%	0.415	0.322	0.458	2.24	2.13	0.848	0.664	0.919	4.45	4.22

Table 1: Results on VAO-Bench. Comp.: within-budget completion rate; mean time averages all registered tasks with timing records, including timeouts. Dir./Der./Inv.: Direct/Derived/Invariant. Bold marks the highest auto-zero score per column; completed-only scores are not ranked because they cover different task subsets.

Figure 4: Difficulty-stratified auto-zero performance of the ten evaluated systems. Horizontal bars show aggregate requirement scores for Easy, Medium, and Hard tasks, with systems ranked independently within each panel. Execution failures receive zero. Best viewed by zoom in .

Condition	Time (s)	Comp.	Auto-zero					Completed-only
Condition	Time (s)	Comp.	Dir.	Der.	Inv.	VQ	DQ	Dir.	Der.	Inv.	VQ	DQ
GPT-Image-2	157.5	100%	0.733	0.551	0.908	4.45	3.83	0.733	0.551	0.908	4.45	3.83
+ Explicit instruction	195.6	99.4%	0.916	0.868	0.928	4.45	4.31	0.923	0.876	0.933	4.48	4.34

Table 2: GPT-Image-2 with explicit instructions on the matched 316-task subset.

A1: image-to-source reconstruction
Condition	Completion	SSIM	PSNR	MAE	MSE	SSIM-0
DeepSeek-Flash A1	92.1%	0.796	16.4	0.0475	0.0276	0.733

Table 3: Native ablations on reconstruction and instruction understanding.

System	Episodes	Completed	Avg. tool calls	Writes artifact	Renders	Views preview
Codex–GPT-5.6-Sol	1573	1525	10.6	1562	1499	1334
Codex–GLM-5.3-Flash	1573	1032	25.9	1572	1550	514
CC–GLM-5.3-Flash	1573	72	29.4	365	343	103

Table 4: Trace statistics. Write, render, and preview counts refer to the designated task artifact, excluding local calibration tests. Despite averaging 29.4 tool calls per episode, CC–GLM-5.3-Flash completes only 72 of 1,573 episodes, compared with 1,032 for Codex–GLM-5.3-Flash and 1,525 for Codex–GPT-5.6-Sol.

Figure 5: Case study of dependency-aware VAO. Changing the objective function requires updating dependent elements while preserving the constraints, region, and layout. Best viewed by zoom in.

Appendix figures & tables32 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Renderer	Sources	Easy	Medium	Hard	Main	Subset
Business webpages	HTML	160	73	67	56	196	40
Scientific documents	HTML	112	49	45	45	139	28
Interactive dashboards	HTML	154	71	62	58	191	38
Flowcharts and workflows	DOT	187	87	78	64	229	46
Systems and architecture	TikZ	107	55	45	34	134	26
Circuits and logic	CircuitikZ	138	68	55	38	161	33

Appendix

Table 5: Composition of VAO-Bench and its ablation subset, VAO-Bench-Ablation. Source counts refer to distinct source artifacts.

Operation	Difficulty	Tasks
Rename labels and text	Easy	200
Change colors, styles, markers, or badges	Easy	128
Edit annotations, captions, or legends	Easy	149
Change highlights, states, or selections	Easy	123
Update local values and synchronize visual elements	Medium	178
Insert or delete local components, reconnect, and reflow	Medium	114

Appendix

Table 6: The 14 operations in VAO-Bench and their task counts.

Field	Meaning and verification
visible_objects , visible_element_ids	Visible text, nodes, values, and graphical objects with unique or identifiable visual anchors.
dependency_graph , data_or_logic_relations	Data, structural, or logical dependencies among visible objects that support linked updates.
visual_recoverability_contract	How labels, tables, or connections provide the evidence needed for an edit.
editable_objects , operation_affordance_matrix	Candidate objects and their supported operations.
stable_regions , stable_invariants	Regions, identities, values, or relations to preserve, extending beyond rectangular pixel regions.

Appendix

Table 7: Construction metadata and its uses. These internal fields are not supplied to solvers in the main experiment.

ID	Method / track	Configuration
M1	GPT-Image-2 / direct	quality=high ; concurrency 5; launch interval 1 s.
M2	Seedream-5.0-Pro / direct	Service ID: doubao-seedream-5-0-pro-260628 1.5K; watermark disabled; concurrency 10; interval 1 s.
M3	GPT-5.6-Sol / native	reasoning_effort=xhigh ; maximum output 65,536 tokens.
M4	DeepSeek-Flash / native	Thinking enabled; reasoning_effort=max ; maximum output 65,536 tokens.
M5	GLM-5.3-Flash / native	reasoning_effort=max ; maximum output 65,536 tokens.
M6	Codex-GPT / agent	Codex + GPT-5.6-Sol; reasoning level medium.

Appendix

Table 8: Evaluated methods and recorded configurations. M1–M10 identify methods in subsequent tables.

Mode	Settings and constraints
HTML	Headless Chromium; 1440 $\times$ 1000 viewport; full-page screenshot; 60 s rendering timeout; self-contained HTML/CSS/JS with no remote resources.
DOT	Graphviz; PNG at 600 DPI; white background; 24 px crop margin; 90 s rendering timeout.
TikZ	pdflatex followed by pdftoppm for PNG conversion; 600 DPI; white background; 96 px crop margin; 90 s timeout; tikz, xcolor, and amsmath allowed.
CircuitikZ	TikZ settings with circuitikz additionally allowed; recorded as a separate rendering mode.
Python	Code runs in the evaluation environment and produces PNG output; allowed packages include Pillow, Matplotlib, NumPy, SciPy, NetworkX, and RDKit.

Appendix

Table 9: Renderers, resolutions, and dependencies used in evaluation.

Difficulty	Tasks	Direct	Derived	Invariant	No Derived
Easy	600	691 / 1.15	9 / 0.01	1546 / 2.58	592
Medium	533	995 / 1.87	884 / 1.66	1594 / 2.99	0
Hard	440	1468 / 3.34	1523 / 3.46	1635 / 3.72	0

Appendix

Table 10: Total scoring points / mean per task. The last column counts tasks without Derived points.

Method	Registered	Normal	Recovered	Timeout	Other
GPT-Image-2	1573	1573	0	0	0
Seedream-5.0-Pro	1573	1478	0	95	0
GPT-5.6-Sol	1573	1529	0	44	0
DeepSeek-Flash	1573	1489	0	84	0
GLM-5.3-Flash	1573	793	0	774	6
Codex-GPT	1573	1525	31	33	15

Appendix

Table 11: Execution coverage. Recovered outputs are a subset of timeouts and are not additional tasks. Other failures include model, code, and rendering errors.

Condition	Inputs	Evaluation and interpretation
Baseline	Source image, original instruction	Main-experiment baseline on the same tasks.
A1	Source image; no editing instruction or ground-truth code	Source reconstruction only; offline pixel metrics and rendering success.
A2	Source image, A1 code, original instruction	Edit the reconstructed representation; standard editing evaluation.
B	Source image, ground-truth source code, original instruction	Control source-representation availability; standard editing evaluation.
C	Source image, expanded instruction; no ground-truth code	Control instruction explicitness; standard editing evaluation.

Appendix

Table 12: Inputs to the DeepSeek-Flash ablations. No editing condition receives the target image.

Figure 6: Semantic strategy actions extracted from execution traces of five agent–model combinations on the same 108 VAO tasks. Each point corresponds to one strategy action at its observed timestamp, with colors denoting action types. The horizontal axis is normalized to the common 30-minute execution budget. Warm colors denote input understanding, reverse engineering, and reconstruction/tuning actions, whereas cool colors denote rendering, comparison, and final answer review or editing.

Figure 7: Cumulative within-budget completion rate over elapsed execution time for five agent–model combinations on the full VAO-Bench evaluation. Each curve shows the fraction of tasks that have completed by the corresponding elapsed time; the 30-minute endpoint corresponds to the reported within-budget completion rate.

Mode	Success/total	SSIM	PSNR	SSIM-0
HTML	106/106	.8147	18.44	.8147
DOT	32/46	.7787	15.39	.5417
TikZ	25/26	.7159	14.33	.6883
CircuitikZ	31/33	.8809	15.63	.8275
Python	97/105	.7740	15.20	.7151

Appendix

Table 13: A1 reconstruction by rendering mode. SSIM and PSNR use successfully rendered outputs; SSIM-0 includes failures.

Panel of Figure 8 — Figure 8: Representative Case 2: source image and reference target (top row), followed by outputs from four rotating methods (middle and bottom rows).

Panel of Figure 9 — Figure 9: Representative Case 3: source image and reference target (top row), followed by outputs from four rotating methods (middle and bottom rows).

Panel of Figure 10 — Figure 10: Representative Case 5: source image and reference target (top row), followed by outputs from four rotating methods (middle and bottom rows).

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	166.2	100.0%	0.689	0.371	0.951	4.821	3.740
M2	266.7	100.0%	0.810	0.590	0.944	4.770	4.010
M3	358.8	100.0%	0.785	0.448	0.971	4.689	4.005
M4	275.7	99.0%	0.795	0.445	0.932	4.740	3.974
M5	1106.4	83.7%	0.640	0.345	0.760	4.036	3.454

Appendix

Table 14: Business webpages: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	177.9	100.0%	0.724	0.515	0.921	4.417	3.583
M2	271.8	100.0%	0.748	0.671	0.939	4.331	3.957
M3	334.2	100.0%	0.784	0.662	0.980	4.345	4.043
M4	364.5	99.3%	0.781	0.649	0.912	4.504	3.978
M5	1257.9	64.7%	0.515	0.416	0.606	2.986	2.640

Appendix

Table 15: Scientific documents: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	160.5	100.0%	0.702	0.381	0.956	4.651	3.413
M2	260.1	100.0%	0.808	0.570	0.958	4.607	3.817
M3	337.8	100.0%	0.740	0.492	0.986	4.539	3.869
M4	342.0	100.0%	0.743	0.495	0.929	4.723	3.885
M5	1087.2	77.0%	0.621	0.455	0.718	3.618	3.204

Appendix

Table 16: Interactive dashboards: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	170.7	100.0%	0.721	0.632	0.898	4.240	3.969
M2	150.3	59.4%	0.478	0.443	0.580	2.642	2.576
M3	362.4	98.3%	0.803	0.724	0.891	3.852	3.974
M4	279.6	96.5%	0.771	0.710	0.815	3.747	3.664
M5	1288.2	44.1%	0.367	0.329	0.394	1.847	1.939

Appendix

Table 17: Flowcharts and workflows: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	156.0	100.0%	0.798	0.628	0.820	4.246	4.246
M2	239.7	100.0%	0.813	0.660	0.960	4.373	4.470
M3	290.4	88.1%	0.786	0.654	0.719	2.448	3.761
M4	511.2	85.8%	0.739	0.612	0.716	3.239	3.657
M5	1610.4	23.9%	0.179	0.165	0.201	0.881	1.060

Appendix

Table 18: Systems and architecture diagrams: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	133.2	100.0%	0.720	0.659	0.952	4.385	3.975
M2	217.5	100.0%	0.781	0.731	0.963	4.435	4.062
M3	221.1	90.7%	0.695	0.659	0.837	3.106	3.689
M4	545.7	84.5%	0.649	0.560	0.743	3.410	3.280
M5	1644.9	18.0%	0.141	0.121	0.160	0.814	0.758

Appendix

Table 19: Circuits and logic diagrams: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	166.8	100.0%	0.815	0.502	0.949	4.258	3.450
M2	250.8	100.0%	0.861	0.648	0.966	4.291	3.907
M3	282.3	99.3%	0.904	0.736	0.981	4.046	4.311
M4	267.6	99.3%	0.911	0.728	0.928	3.861	4.066
M5	1435.2	50.3%	0.457	0.347	0.487	2.086	2.093

Appendix

Table 20: Scientific statistical charts: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	138.9	100.0%	0.664	0.390	0.877	4.263	3.359
M2	218.4	100.0%	0.794	0.641	0.912	4.174	3.701
M3	277.2	95.8%	0.918	0.896	0.942	4.245	4.677
M4	343.5	95.2%	0.876	0.876	0.817	3.826	4.198
M5	1373.1	46.1%	0.394	0.351	0.410	1.964	2.084

Appendix

Table 21: Mathematics and geometry: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	120.3	100.0%	0.770	0.706	0.903	4.785	4.278
M2	199.2	99.0%	0.809	0.752	0.927	4.566	4.258
M3	216.3	99.5%	0.850	0.752	0.954	4.688	4.488
M4	343.5	90.2%	0.745	0.636	0.650	3.668	3.615
M5	1490.1	37.6%	0.306	0.280	0.292	1.532	1.566

Appendix

Table 22: Chemical diagrams: auto-zero and completed-only results.

Method	Population	Direct	Derived	Invariant
GPT-Image-2	Auto-zero	3154	2416	4775
GPT-Image-2	Completed-only	3154	2416	4775
Seedream-5.0-Pro	Auto-zero	3154	2416	4775
Seedream-5.0-Pro	Completed-only	2967	2275	4477
GPT-5.6-Sol	Auto-zero	3154	2416	4775
GPT-5.6-Sol	Completed-only	3060	2348	4643

Appendix

Table 23: Overall scoring-point denominators. Auto-zero includes execution failures; completed-only includes successfully scored tasks. Both include successfully scored outputs recovered after timeout.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	150.3	100.0%	0.897	0.556	0.929	4.570	4.520
M2	210.0	93.3%	0.884	0.667	0.900	4.300	4.338
M3	307.2	97.0%	0.916	0.556	0.920	4.115	4.368
M4	339.3	95.3%	0.909	0.556	0.810	4.027	4.097
M5	1304.7	52.0%	0.501	0.556	0.457	2.327	2.363

Appendix

Table 24: Easy tasks: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	153.3	100.0%	0.772	0.546	0.925	4.413	3.616
M2	224.7	94.7%	0.777	0.640	0.898	4.176	3.764
M3	296.7	97.7%	0.846	0.656	0.933	4.036	4.002
M4	358.8	94.9%	0.826	0.619	0.833	3.993	3.707
M5	1348.2	52.2%	0.457	0.336	0.487	2.330	2.161

Appendix

Table 25: Medium tasks: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	160.2	100.0%	0.622	0.515	0.895	4.396	3.000
M2	251.7	93.9%	0.682	0.611	0.872	4.084	3.136
M3	294.6	96.8%	0.728	0.668	0.921	4.002	3.845
M4	372.0	93.4%	0.683	0.639	0.835	3.927	3.530
M5	1420.5	46.1%	0.345	0.313	0.429	2.018	1.770

Appendix

Table 26: Hard tasks: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	167.4	100.0%	0.703	0.414	0.944	4.653	3.580
M2	265.8	100.0%	0.793	0.604	0.948	4.595	3.926
M3	344.7	100.0%	0.768	0.522	0.979	4.544	3.966
M4	323.1	99.4%	0.772	0.519	0.925	4.671	3.943
M5	1139.4	76.2%	0.600	0.405	0.703	3.607	3.148

Appendix

Table 27: HTML rendering: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	170.7	100.0%	0.721	0.632	0.898	4.240	3.969
M2	150.3	59.4%	0.478	0.443	0.580	2.642	2.576
M3	362.4	98.3%	0.803	0.724	0.891	3.852	3.974
M4	279.6	96.5%	0.771	0.710	0.815	3.747	3.664
M5	1288.2	44.1%	0.367	0.329	0.394	1.847	1.939

Appendix

Table 28: DOT rendering: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	156.0	100.0%	0.798	0.628	0.820	4.246	4.246
M2	239.7	100.0%	0.813	0.660	0.960	4.373	4.470
M3	290.4	88.1%	0.786	0.654	0.719	2.448	3.761
M4	511.2	85.8%	0.739	0.612	0.716	3.239	3.657
M5	1610.4	23.9%	0.179	0.165	0.201	0.881	1.060

Appendix

Table 29: TikZ rendering: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	133.2	100.0%	0.720	0.659	0.952	4.385	3.975
M2	217.5	100.0%	0.781	0.731	0.963	4.435	4.062
M3	221.1	90.7%	0.695	0.659	0.837	3.106	3.689
M4	545.7	84.5%	0.649	0.560	0.743	3.410	3.280
M5	1644.9	18.0%	0.141	0.121	0.160	0.814	0.758

Appendix

Table 30: CircuitikZ rendering: auto-zero and completed-only results.

Auto-zero
ID	Time (s)	Completion	Dir.	Der.	Inv.	VQ	DQ
M1	139.8	100.0%	0.749	0.539	0.908	4.466	3.746
M2	220.2	99.6%	0.819	0.684	0.934	4.361	3.979
M3	254.7	98.3%	0.888	0.795	0.958	4.361	4.497
M4	321.6	94.5%	0.836	0.744	0.782	3.774	3.931
M5	1437.0	44.0%	0.379	0.324	0.385	1.830	1.883

Appendix

Table 31: Python rendering: auto-zero and completed-only results.

Image Editing Evaluation

Image editing evaluation measures whether edits follow instructions while preserving identity, context, and visual quality.

See all 92 papers

PaintBench: Deterministic Evaluation of Precise Visual Editing
May 29, 2026Kai Xu, Ellis Brown, Shrikar Madhu +3Image Editing Evaluation Text-Guided Image Editing
New York University
VisEditBench: Can Vision-Language Models Edit Visualization Code from Multimodal Feedback?
Aug 11, 2026Mizanur Rahman, Arshia Azimlu, Shadikur Rahman +4VLM Evaluation Data Visualization
‡York University · $Nanyang Technological University · ¶Salesforce AI Research
VDE Bench: Evaluating The Capability of Image Editing Models to Modify Visual Documents
Jan 27, 2026Hongzhu Yi, Yujia Yang, Yuanxiang Wang +18Image Editing Evaluation Image Editing
UCAS · Coserve AI · WashU +5

May 29, 2026J/K move · Enter open · S save

PaintBench: Deterministic Evaluation of Precise Visual Editing

Kai Xu, Ellis Brown, Shrikar Madhu, Rob Fergus+2

New York University

Image Editing Evaluation Text-Guided Image Editing Image Editing

Open paper AskPDF arXiv

While current multimodal models are proficient at open-ended visual editing, executing precise single-answer edits remains an important obstacle. To probe this challenge, we introduce PaintBench, a dynamically scalable benchmark targeting 20 fundamental precise visual editing operations across four categories: geometric transformation, structural manipulation, color change, and symbolic reasoning. Procedural generation with configurable complexity enables an effectively infinite, contamination-resistant evaluation suite, and deterministic pixel-level evaluation eliminates reliance on bias-prone judge models. Across 11 image editing models, we find overall low performance, with the current highest-performing industry leader scoring only 17.1% (mIoU). Task decomposition reveals especially challenging operation types (geometric transformation, most structural manipulation, formula-based color change) and model-specific specializations. Fine-grained benchmark diagnostics further show performance degradations induced by scene variations in object count, background complexity, color scheme, and edit-region size. To test generalization of PaintBench scores to applied task performance, we create a procedural, deterministic evaluation for data visualization editing (TinyGrafixBench) and find strong linear correlation with PaintBench scores (

R^2 = 0.91

p < 0.001

). Altogether, PaintBench provides a rigorous foundation for measuring and driving progress in precise multimodal visual editing.

Abstract

Image Editing Evaluation

Explore similar work

PaintBench: Deterministic Evaluation of Precise Visual Editing

VisEditBench: Can Vision-Language Models Edit Visualization Code from Multimodal Feedback?

VDE Bench: Evaluating The Capability of Image Editing Models to Modify Visual Documents