Writing study: PhotoAgent: Exploratory Visual Aesthetic Planning with Large Vision Models

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Mingde Yao, Zhiyuan You, King Man Tam, Menglu Wang, Tianfan Xue · Official ICML 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Domain-grounded methods & physical systems (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

31 physical pages

Before references / appendix

≈ 8.71 page-equivalents

Main figures / tables

6 / 2

Attached appendix figures / tables

7 / 4

Attached appendix span

≈ 18.91 page-equivalents

Before references / appendix: ≈ 28.1%References / intervening material: ≈ 10.9%Attached appendix: ≈ 61%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1. Introduction	1	≈ 1.93 pp.	6.23%
2. Related Work	3	≈ 1.09 pp.	3.52%
3. PhotoAgent	4	≈ 1.36 pp.	4.39%
4. Evaluation for Editing Systems	5	≈ 0.96 pp.	3.1%
5. Experiments	6	≈ 2.65 pp.	8.55%
6. Conclusion	9	≈ 0.21 pp.	0.68%
Acknowledgements	9	≈ 0.07 pp.	0.23%
Impact Statement	9	≈ 0.1 pp.	0.32%
References	9	≈ 3.38 pp.	10.9%
A. Sim-to-Real Gap in Low-Resolution Planner Simulation	13	≈ 0.43 pp.	1.39%
B. Generalization of UGC Reward Model	13	≈ 0.23 pp.	0.74%
C. Experimental details	13	≈ 6.43 pp.	20.74%
D. More Visual Results	20	≈ 0.63 pp.	2.03%
E. Dataset Diversity and Fairness	20	≈ 0.44 pp.	1.42%
F. Computational Cost	21	≈ 1.02 pp.	3.29%
G. Future Work	22	≈ 0.46 pp.	1.48%
H. Prompt	22	≈ 9.26 pp.	29.87%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 4
• Figure 3 · p. 6
• Figure 4 · p. 7
• Table 1 · p. 7
• Table 2 · p. 8
• Figure 5 · p. 8
• Figure 6 · p. 9

Attached appendix

• Table 3 · p. 13
• Table 4 · p. 13
• Figure 7 · p. 14
• Figure 8 · p. 15
• Figure 9 · p. 16
• Figure 10 · p. 17
• Figure 11 · p. 18
• Figure 12 · p. 19
• Table 5 · p. 21
• Table 6 · p. 21
• Figure 13 · p. 22

Argument and organization

1. Problem and contribution

   Connects expertise, tool selection and interaction burdens to a perceive-plan-execute-evaluate loop that can propose semantic edits without detailed user commands.

2. Argument sequence

   Each component receives its own method paragraph; the evaluation resource is separately introduced before end-to-end comparisons and planning ablations.

3. Evidence and verification

   Human preference uses randomized unlabeled forced choice, while the supplement tests low-resolution/full-resolution ranking agreement and an external reward-model dataset.

Problem, claim and contribution

Connects expertise, tool selection and interaction burdens to a perceive-plan-execute-evaluate loop that can propose semantic edits without detailed user commands.

Introduction and Figure 1, p. 2 · Main paper

Argument and section order

Each component receives its own method paragraph; the evaluation resource is separately introduced before end-to-end comparisons and planning ablations.

Sections 3-5, pp. 4-8 · Main paper

Figures, captions and visual explanation

The circular overview makes rollback explicit; the comparison table retains metrics won by another model, and adjacent iteration panels show the sequence rather than only the final edit.

Figure 1, p. 2; Tables 1-2 and Figure 5, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

Human preference uses randomized unlabeled forced choice, while the supplement tests low-resolution/full-resolution ranking agreement and an external reward-model dataset.

User Study, p. 8; Tables 3-4, p. 13 · Main paper

Appendix and reproducibility

Modular prompts distinguish output constraints, image context, scene-specific aesthetics, history and user requirements, exposing the actual policy interface.

Appendix H, pp. 22-31 · Main paper

Clarity and writing conventions

Concrete editing trajectories help explain the workflow, but qualitative executor comparisons do not alone isolate planning as the cause of gains.

Sections 5.2-5.4, p. 8 · Main paper

What to reuse—and what not to infer

Transferable lesson: For an agent application, document rollback, approximation checks and the full prompt interface alongside aggregate results.

Boundary: Aesthetic preference is population- and task-dependent; full-resolution rescoring cannot guarantee recovery of a good candidate already omitted by low-resolution search.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-6, pp. 1-9: editing-loop motivation, four-component planner, reward dataset, comparisons, human preference and ablations.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-H, pp. 13-31: low/full-resolution alignment, reward transfer, planner details, examples, costs, future work and full modular prompts.

Visually inspected PDF pages

2, 8

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 31 pages