PaperFigure

Open retrieval. Evidence-grounded. Main figure only.

An original Codex skill for turning AI research into clear, publication-ready main figures.

PaperFigure researches the target method, retrieves candidate papers, reads their actual Figure images together with captions and method context, ranks visual structure at Figure level, plans an original visual explanation, renders one main figure, and critiques it for factual fidelity, exact text, and readability.

It is intentionally narrow: method overviews, architectures, training/inference pipelines, and conceptual frameworks. It does not generate result plots, posters, slide decks, or a full paper illustration suite.

Why PaperFigure

Open scholarly retrieval - searches beyond a fixed reference bank, prioritizing arXiv and official proceedings such as NeurIPS, ICML, ICLR, ACL, EMNLP, CVPR, ICCV, ECCV, AAAI, and KDD.
Actual Figure understanding - extracts Figure images, captions, and surrounding method text; identifies panels, layout, modules, arrows, loops, branching, hierarchy, insets, and frozen/trainable encoding.
Figure-level multimodal ranking - ranks each Figure by topic (35%), visual structure (35%), multi-label intent (15%), source quality (10%), and recency (5%).
Explicit fallback provenance - image+text means the image was inspected; unavailable images are marked text-only and receive reduced structure confidence.
Evidence/appearance separation - the primary paper determines facts; retrieved figures influence visual grammar only.
Main-figure discipline - extracts the smallest set of components needed to communicate the paper's central contribution.
Four-gate critique - checks factual fidelity, character-level text correctness, thumbnail-scale readability, and semantic palette discipline.
No API key for retrieval - the included Python search utility uses public arXiv and OpenAlex endpoints and only the standard library.

Workflow

Paper / method notes
        |
        v
Target Visual Profile       Open literature paper recall
        |                              |
        |                    Candidate Figure extraction
        |                              |
        |              image + caption + method context
        |                              |
        +-------> VisualFigureProfile ranking <---------+
                         |
                         v
             Evidence-grounded Figure Brief
                         |
                         v
                  Main-figure generation
                         |
                         v
       Fidelity -> Text -> Readability -> Palette gates
                         |
                         v
        main-figure.png + brief.md + evidence.json

Install

git clone https://github.com/goya4140/PaperFigure.git
cp -R PaperFigure/paper-figure ~/.codex/skills/paper-figure

The default ar5iv extraction path uses only Python's standard library. Install the PDF fallback when needed:

python3 -m pip install -r PaperFigure/requirements.txt

Restart Codex after installation, then invoke the skill directly or let Codex select it automatically.

Use

Use $paper-figure to turn this paper into a publication-ready main figure:
https://arxiv.org/abs/2510.24940

You can provide an arXiv/proceedings URL, title, abstract, method section, draft, or structured notes. The standard deliverables are:

main-figure.png
brief.md
evidence.json
target-visual-profile.json
candidate VisualFigureProfile files
visual-ranking.json

For reproducible candidate discovery:

python3 paper-figure/scripts/search_open_literature.py \
  "open vocabulary segmentation language prompts SAM2" \
  --limit 8 \
  --output retrieval.json

Paper recall intentionally assigns no visual score. Continue with Figure extraction and multimodal profiling:

python3 paper-figure/scripts/figure_extractor.py 2602.08224 \
  --output-dir work/efficient-sam2 --max-figures 3

python3 paper-figure/scripts/visual_profile.py build-requests \
  work/efficient-sam2/figures.json \
  --paper-title "Efficient-SAM2" \
  --topic-text "efficient SAM2 segmentation" \
  --output work/profile-requests.json

# Inspect each image with a vision-capable model, save profile JSON, then:
python3 paper-figure/scripts/visual_profile.py validate work/profiles/*.json
python3 paper-figure/scripts/visual_ranker.py \
  --target target-visual-profile.json \
  --profiles work/profiles/*.json \
  --output visual-ranking.json

Supported multi-label intents are architecture, pipeline, training-inference, iterative-optimization, multi-agent, data-flow, and conceptual-comparison.

Multimodal ranking verification

The repository includes a real image-level verification run over five candidate Figures. Three method Figures were ranked for SemCoT, OpenWorldSAM, and DataRater; two attention/dataset grids were excluded after image inspection.

View inspected sources, profiles, score breakdowns, and results

Palette-optimized benchmark

PaperFigure now plans color as a semantic system instead of a collection of decorative hues: a pure-white publication canvas, one dark ink, pale role tints, no more than three chromatic roles plus neutrals, and an approximate 80/15/5 area balance that reserves saturated accent color for the paper's central intervention.

The complete ten-case gallery below now uses this protocol. The original 4DGS-1K regression, including its brief, pixel audit, and Gemini revision trace, remains available here.

Ten-case Codex vs Gemini benchmark

The benchmark contains ten PaperBananaBench diagram cases. The completed 4DGS-1K palette regression is retained, while the other nine cases were rerun for this update. For every case, the same palette-optimized prompt was supplied to the Codex image-generation path and Google's gemini-3.1-flash-image; narrow issue-specific revisions are noted, and unresolved failures remain visible.

Every image was inspected at original resolution against PaperFigure's four gates: factual fidelity, exact text, visual readability, and palette/white background. Codex passes all four gates on 10/10 cases and Gemini on 6/10; all 20 images pass the readability and palette gates. Open the full protocol, shared prompts, and structured results.

#	Case	Source	Codex	Gemini
1	SemCoT	`test.json:test_233`	PASS	PASS
2	OpenWorldSAM	`ref.json:ref_191`	PASS	PASS
3	DataRater	diagram image collection	PASS	REJECTED
4	3D-Prover	`test.json:test_2`	PASS	REJECTED
5	4DGS-1K	`test.json:test_1`	PASS	PASS
6	MINI	`test.json:test_5`	PASS	REJECTED
7	DyFlow	`test.json:test_68`	PASS	PASS
8	AutoData	`ref.json:ref_26`	PASS	REJECTED
9	HyperGraphRAG	`ref.json:ref_123`	PASS	PASS
10	BIFROST-1	`ref.json:ref_34`	PASS	PASS

1. SemCoT

Codex	Gemini 3.1 Flash Image

PASS _{Factual · Exact text · Readability · Palette}	PASS _{All four gates after targeted revision}

2. OpenWorldSAM

Codex	Gemini 3.1 Flash Image

PASS _{Factual · Exact text · Readability · Palette}	PASS _{Factual · Exact text · Readability · Palette}

3. DataRater

Codex	Gemini 3.1 Flash Image

PASS _{Factual · Exact text · Readability · Palette}	REJECTED _{Literal theta text still fails after revision}

4. 3D-Prover

Codex	Gemini 3.1 Flash Image

PASS _{All four gates after targeted revision}	REJECTED _{Predictor output topology still fails}

5. 4DGS-1K

Codex	Gemini 3.1 Flash Image

PASS _{Factual · Exact text · Readability · Palette}	PASS _{Factual · Exact text · Readability · Palette}

6. MINI

Codex	Gemini 3.1 Flash Image

PASS _{All four gates after white-background revision}	REJECTED _{Duplicated backbone regions remain}

7. DyFlow

Codex	Gemini 3.1 Flash Image

PASS _{Factual · Exact text · Readability · Palette}	PASS _{All four gates after targeted revision}

8. AutoData

Codex	Gemini 3.1 Flash Image

PASS _{Factual · Exact text · Readability · Palette}	REJECTED _{Duplicated external Web endpoint remains}

9. HyperGraphRAG

Codex	Gemini 3.1 Flash Image

PASS _{All four gates after white-background revision}	PASS _{Factual · Exact text · Readability · Palette}

10. BIFROST-1

Codex	Gemini 3.1 Flash Image

PASS _{Factual · Exact text · Readability · Palette}	PASS _{All four gates after targeted revision}

Repository layout

PaperFigure/
├── paper-figure/
│   ├── SKILL.md                 # Codex workflow and boundaries
│   ├── agents/openai.yaml       # Skill UI metadata
│   ├── references/              # Retrieval, profile, and critique protocols
│   └── scripts/                 # Paper recall, Figure extraction/profile/ranking
├── examples/                    # Benchmark runs and multimodal ranking evidence
├── tests/                       # Retrieval/ranking unit tests
├── LICENSE
└── README.md

Validate

python3 -m unittest discover -s tests -v
python3 paper-figure/scripts/visual_profile.py validate examples/multimodal-ranking/profiles/*.json
python3 /path/to/skill-creator/scripts/quick_validate.py paper-figure

Design principles

The target paper is the authority for content.
Retrieved papers are references for visual structure, never a license to copy a figure.
Venue priority breaks ties; it does not rescue a weak topic or intent match.
A concise main figure is better than a complete but unreadable method dump.
If the method cannot be verified, PaperFigure stops at a draft brief instead of rendering false certainty.

License

MIT

Navigation Menu

Repository files navigation

PaperFigure

Open retrieval. Evidence-grounded. Main figure only.

Why PaperFigure

Workflow

Install

Use

Multimodal ranking verification

Palette-optimized benchmark

Ten-case Codex vs Gemini benchmark

1. SemCoT

2. OpenWorldSAM

3. DataRater

4. 3D-Prover

5. 4DGS-1K

6. MINI

7. DyFlow

8. AutoData

9. HyperGraphRAG

10. BIFROST-1

Repository layout

Validate

Design principles

License

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
examples		examples
paper-figure		paper-figure
tests		tests
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation

PaperFigure

Open retrieval. Evidence-grounded. Main figure only.

Why PaperFigure

Workflow

Install

Use

Multimodal ranking verification

Palette-optimized benchmark

Ten-case Codex vs Gemini benchmark

1. SemCoT

2. OpenWorldSAM

3. DataRater

4. 3D-Prover

5. 4DGS-1K

6. MINI

7. DyFlow

8. AutoData

9. HyperGraphRAG

10. BIFROST-1

Repository layout

Validate

Design principles

License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages