PaperBanana Automated Academic Illustration for AI Scientists

---

Disclaimer: This is an unofficial, community-driven open-source implementation of the paper "PaperBanana: Automating Academic Illustration for AI Scientists" by Dawei Zhu, Rui Meng, Yale Song, Xiyu Wei, Sujian Li, Tomas Pfister, and Jinsung Yoon (arXiv:2601.23265). This project is not affiliated with or endorsed by the original authors or Google Research. The implementation is based on the publicly available paper and may differ from the original system.

An agentic framework for generating publication-quality academic diagrams and statistical plots from text descriptions. Supports OpenAI (GPT-5.2 + GPT-Image-1.5), Azure OpenAI / Foundry, Google Gemini, and Atlas Cloud providers.

• Two-phase multi-agent pipeline with iterative refinement
• Multiple VLM and image generation providers (OpenAI, Azure, Gemini, Atlas Cloud)
• Input optimization layer for better generation quality
• Auto-refine mode and run continuation with user feedback
• CLI, Python API, and MCP server for IDE integration
• Batch generation from a manifest file (YAML/JSON) for multiple diagrams in one run
• Batch plots — paperbanana plot-batch runs many statistical plots from one manifest (CSV/JSON per item)
• PDF inputs for methodology context (optional paperbanana[pdf] / PyMuPDF), with per-page selection
• PaperBanana Studio — local Gradio web UI (paperbanana studio) for diagrams, plots, evaluation, batch, and run browser
• Claude Code skills for /generate-diagram, /generate-plot, and /evaluate-diagram

Atlas Cloud

Atlas Cloud is a full-modal AI inference platform that gives developers a single AI API to access video generation, image generation, and LLM APIs. Instead of managing multiple vendor integrations, you connect once and get unified access to 300+ curated models across all modalities.

Check out Atlas Cloud's new coding plan promotion for more budget-friendly API access: https://www.atlascloud.ai/console/coding-plan

---

Quick Start

Try it in your browser: the Colab quickstart notebook walks through install → API key → diagram generation end-to-end, no local setup required.

Prerequisites

• Python 3.10+
• An OpenAI API key (platform.openai.com) or Azure OpenAI / Foundry endpoint
• Or a Google Gemini API key (free, Google AI Studio)

Step 1: Install

pip install paperbanana

Or install from source for development:

git clone https://github.com/llmsresearch/paperbanana.git
cd paperbanana
pip install -e ".[dev,openai,google]"

Docker

Build the image from a clone of the repo and pass your API key at runtime:

docker build -t paperbanana .
docker run --rm -e GOOGLE_API_KEY paperbanana generate --help

To generate a diagram, mount your input and an outputs folder into /work:

docker run --rm -e GOOGLE_API_KEY \
  -v "$(pwd)/method.txt:/work/method.txt:ro" \
  -v "$(pwd)/outputs:/work/outputs" \
  paperbanana generate --input method.txt --caption "Overview of our framework"

Step 2: Get Your API Key

cp .env.example .env
# Edit .env and add your API key:
#   OPENAI_API_KEY=your-key-here
#   GOOGLE_API_KEY=your-key-here
#
# For Azure OpenAI / Foundry:
#   OPENAI_BASE_URL=https://<resource>.openai.azure.com/openai/v1
#
# Optional Gemini overrides:
#   GOOGLE_BASE_URL=https://your-gemini-proxy.example.com
#   GOOGLE_VLM_MODEL=gemini-2.5-flash
#   GOOGLE_IMAGE_MODEL=gemini-3-pro-image-preview

Or use the setup wizard for Gemini:

paperbanana setup

Step 3: Generate a Diagram

paperbanana generate \
  --input examples/sample_inputs/transformer_method.txt \
  --caption "Overview of our encoder-decoder architecture with sparse routing"

With input optimization and auto-refine:

paperbanana generate \
  --input my_method.txt \
  --caption "Overview of our encoder-decoder framework" \
  --optimize --auto

Output is saved to outputs/run_<timestamp>/final_output.png along with all intermediate iterations and metadata.

PaperBanana Studio (local web UI)

Install the optional Gradio dependency, then start the app:

pip install 'paperbanana[studio]'
paperbanana studio

Open the URL shown in the terminal (default http://127.0.0.1:7860/). The Studio exposes the same workflows as the CLI: methodology diagrams, statistical plots, comparative evaluation, continuing a prior run, batch manifests (methodology or plot batch via the Batch tab), and a simple browser for run_* / batch_* output folders. Use --host, --port, --config, and --output-dir as needed.

---

How It Works

PaperBanana implements a multi-agent pipeline with up to 7 specialized agents:

Phase 0 -- Input Optimization (optional, --optimize):

0. Input Optimizer runs two parallel VLM calls:
   • Context Enricher structures raw methodology text into diagram-ready format (components, flows, groupings, I/O)
   • Caption Sharpener transforms vague captions into precise visual specifications

Phase 1 -- Linear Planning:

1. Retriever selects the most relevant reference examples from a curated set of 13 methodology diagrams spanning agent/reasoning, vision/perception, generative/learning, and science/applications domains
2. Planner generates a detailed textual description of the target diagram via in-context learning from the retrieved examples
3. Stylist refines the description for visual aesthetics using NeurIPS-style guidelines (color palette, layout, typography)

Phase 2 -- Iterative Refinement:

4. Visualizer renders the description into an image
5. Critic evaluates the generated image against the source context and provides a revised description addressing any issues
6. Steps 4-5 repeat for a fixed number of iterations (default 3), or until the critic is satisfied (--auto)

Providers

PaperBanana supports multiple VLM and image generation providers:

Component	Provider	Model	Notes
VLM (planning, critique)	OpenAI	gpt-5.2	Default
Image Generation	OpenAI	gpt-image-1.5	Default
VLM	Atlas Cloud	deepseek-ai/DeepSeek-V3-0324	OpenAI-compatible chat endpoint
Image Generation	Atlas Cloud	openai/gpt-image-2/text-to-image	Async prediction API
VLM	Google Gemini	gemini-2.5-flash	Low cost
Image Generation	Google Gemini	gemini-3-pro-image-preview	$0.134/image (1K)
VLM / Image	OpenRouter	Any supported model	Flexible routing

Azure OpenAI / Foundry endpoints are auto-detected — set OPENAI_BASE_URL to your endpoint. Gemini-compatible gateways are also supported — set GOOGLE_BASE_URL when needed. Atlas Cloud uses ATLASCLOUD_BASE_URL=https://api.atlascloud.ai/v1 for chat and ATLASCLOUD_IMAGE_BASE_URL=https://api.atlascloud.ai/api/v1 for image generation.

Atlas Cloud official site: https://www.atlascloud.ai/?utm_source=github&utm_medium=link&utm_campaign=paperbanana

Recommended Atlas LLM models for ATLASCLOUD_VLM_MODEL:

• deepseek-ai/DeepSeek-V3-0324 (default)
• openai/gpt-4o
• openai/gpt-4.1
• google/gemini-2.5-flash
• anthropic/claude-sonnet-4.5-20250929

These are stable, generally available models verified against the Atlas Cloud API. The full, always-current model pool (300+ models) is documented on Atlas Cloud's own docs — see https://www.atlascloud.ai/models — and any model id listed there can be passed via ATLASCLOUD_VLM_MODEL.

Recommended Atlas image models for ATLASCLOUD_IMAGE_MODEL:

• openai/gpt-image-2/text-to-image
• openai/gpt-image-2/edit
• baidu/ERNIE-Image-Turbo/text-to-image
• black-forest-labs/flux-dev
• black-forest-labs/flux-schnell
• qwen/qwen-image

---

CLI Reference

paperbanana generate -- Methodology Diagrams

# Basic generation
paperbanana generate \
  --input method.txt \
  --caption "Overview of our framework"

# With input optimization and auto-refine
paperbanana generate \
  --input method.txt \
  --caption "Overview of our framework" \
  --optimize --auto

# Continue the latest run with user feedback
paperbanana generate --continue \
  --feedback "Make arrows thicker and colors more distinct"

# Continue a specific run
paperbanana generate --continue-run run_20260218_125448_e7b876 \
  --iterations 3

# PDF as input (install PyMuPDF: pip install 'paperbanana[pdf]')
paperbanana generate \
  --input paper.pdf \
  --caption "Overview of our method" \
  --pdf-pages "3-8"

# Guide generation with a reference/sketch image (repeatable)
paperbanana generate \
  --input method.txt \
  --caption "Overview of our framework" \
  --image sketch.png --image prior_figure.png

Flag	Short	Description
--input	-i	Path to methodology text file or PDF (required for new runs)
--caption	-c	Figure caption / communicative intent (required for new runs)
--image		Reference/sketch image (hand-drawn sketch, whiteboard photo, prior figure) that guides the Planner. Repeatable for multiple images
--output	-o	Output image path (default: auto-generated in outputs/)
--iterations	-n	Number of Visualizer-Critic refinement rounds (default: 3)
--num-candidates	-k	Generate N candidate images in parallel, 1-8 (default: 1). Planning runs once; refinement fans out per candidate with seed offsets. Outputs land in candidates/cand_<i>/; the run-root final_output is candidate 1. Cost estimates and --budget account for the fan-out
--auto		Loop until critic is satisfied (with --max-iterations safety cap)
--max-iterations		Safety cap for --auto mode (default: 30)
--optimize		Preprocess inputs with parallel context enrichment and caption sharpening
--continue		Continue from the latest run in outputs/
--continue-run		Continue from a specific run ID
--feedback		User feedback for the critic when continuing a run
--pdf-pages		PDF input only: 1-based pages (e.g. 1-5, 2,4,6-8; default: all)
--vlm-provider		VLM provider name (default: openai)
--vlm-model		VLM model name (default: gpt-5.2)
--image-provider		Image gen provider (default: openai_imagen)
--image-model		Image gen model (default: gpt-image-1.5)
--format	-f	Output format: png, jpeg, or webp (default: png)
--config		Path to YAML config file (see configs/config.yaml)
--verbose	-v	Show detailed agent progress and timing
--progress-json		Emit JSON progress events to stdout during generation

paperbanana plot -- Statistical Plots

paperbanana plot \
  --data results.csv \
  --intent "Bar chart comparing model accuracy across benchmarks"

Flag	Short	Description
--data	-d	Path to data file, CSV or JSON (required)
--intent		Communicative intent for the plot (required)
--output	-o	Output image path
--iterations	-n	Refinement iterations (default: 3)
--vlm-provider		VLM provider name
--vlm-model		VLM model name

Plots are rendered via VLM-generated matplotlib code — no image-generation provider or credentials are required.

paperbanana venues -- Custom Venue Style Packs

--venue selects a venue style pack: a directory with methodology_style_guide.md, plot_style_guide.md, and an optional venue.yaml. Built-in packs (neurips, icml, acl, ieee) ship with PaperBanana; you can add your own under ~/.config/paperbanana/venues/ (override with --venue-dir or PAPERBANANA_VENUE_DIR) without touching the repo:

# 1. Scaffold a pack (seeds both guides from the NeurIPS templates)
paperbanana venues init mylab

# 2. Edit the style guides — or generate them from a corpus of example figures:
#    paperbanana guidelines synthesize --reference-set ./examples \
#      --output ~/.config/paperbanana/venues/mylab/methodology_style_guide.md

# 3. Use it anywhere --venue is accepted
paperbanana generate --input method.txt --caption "Overview" --venue mylab

# See everything that's available (built-in + user, with source)
paperbanana venues list

venue.yaml (all fields optional):

display_name: "My Lab Style"  # shown by `paperbanana venues list`
aspect_ratio: "16:9"          # default --aspect-ratio for this venue's runs
fonts:                        # preferred fonts, appended to the style guides
  - "Helvetica"

On a name clash, built-in packs win — user packs cannot shadow built-in venues. Unknown venue names fail fast with the list of available packs from both sources.

paperbanana batch -- Batch Generation

Generate multiple methodology diagrams from a single manifest file (YAML or JSON). Each item runs the full pipeline; outputs are written under outputs/batch_<id>/run_<id>/ and a batch_report.json summarizes all runs.

paperbanana batch --manifest examples/batch_manifest.yaml --optimize

Manifest format (YAML or JSON with an items list):

items:
  - input: path/to/method1.txt
    caption: "Overview of our encoder-decoder"
    id: fig1
  - input: method2.txt
    caption: "Training pipeline"
    id: fig2
  - input: paper.pdf
    caption: "System overview"
    id: fig3
    pdf_pages: "4-9" # optional; PDF inputs only

Paths in the manifest are resolved relative to the manifest file's directory.

Composite figures: Add an optional composite section to automatically stitch all generated panels into a single labeled figure after the batch completes:

composite:
  layout: "1x3"          # rows x cols, or "auto"
  labels: auto            # (a), (b), (c)... or explicit list, or null
  spacing: 20             # pixels between panels
  label_position: bottom  # top or bottom
  output: "composite.png"

items:
  - input: method_encoder.txt
    caption: "Encoder architecture"
    id: panel_a
  # ...

The composite image is saved alongside the individual panels in the batch output directory. See examples/composite_batch_manifest.yaml for a complete example.

Generate a human-readable report from an existing batch run (Markdown or HTML):

paperbanana batch-report --batch-dir outputs/batch_20250109_123456_abc --format markdown
# or by batch ID (under default output dir)
paperbanana batch-report --batch-id batch_20250109_123456_abc --format html --output report.html

Diagram batch reports include batch_kind: methodology; plot batches use batch_kind: statistical_plot. Human-readable reports (paperbanana batch-report) show the batch kind when present.

Sweep manifests let you store the full sweep plan as YAML/JSON instead of eight comma-separated CLI flags. Mutually exclusive with the axis flags; see examples/sweep_manifest.yaml.

paperbanana sweep --manifest examples/sweep_manifest.yaml

Sweep reports produced by paperbanana sweep can be rendered the same way:

paperbanana sweep-report --sweep-dir outputs/sweep_20250109_123456_abc --format html
# or by sweep ID
paperbanana sweep-report --sweep-id sweep_20250109_123456_abc --format markdown

Rendered sweep reports include a summary, a top-5 ranked table, the full variants table (with per-variant provider/model, iterations, critic-suggestion count, proxy score, and output path), and the quality_proxy_score note. Dry-run reports render a simplified "Planned Variants" section.

Flag	Short	Description
--manifest	-m	Path to manifest file (required)
--output-dir	-o	Parent directory for batch run (default: outputs)
--config		Path to config YAML
--iterations	-n	Refinement iterations per item
--optimize		Preprocess inputs for each item
--auto		Loop until critic satisfied per item
--format	-f	Output image format (png, jpeg, webp)
--auto-download-data		Auto-download the PaperBananaBench reference set (~254 MB) if not cached

paperbanana plot-batch -- Batch Statistical Plots

Generate multiple plots from a manifest (YAML or JSON). Each item specifies a data file (CSV or JSON) and an intent string, mirroring paperbanana plot. Outputs live under outputs/batch_<id>/run_<id>/ with the same batch_report.json and paperbanana batch-report workflow as diagram batches.

paperbanana plot-batch --manifest examples/plot_batch_manifest.yaml --optimize

Manifest format (items list):

items:
  - data: path/to/results.csv
    intent: "Bar chart comparing accuracy across models"
    id: fig_acc
  - data: other.json
    intent: "Scatter plot with trend line"
    aspect_ratio: "16:9"   # optional per item; CLI --aspect-ratio is the default when omitted

Paths are resolved relative to the manifest file’s directory.

Flag	Short	Description
--manifest	-m	Path to manifest (required)
--output-dir	-o	Parent directory for batch_* (default: outputs)
--config		Path to config YAML
--vlm-provider		VLM provider (default: gemini)
--vlm-model		VLM model override
--image-provider		Image gen provider
--image-model		Image gen model
--iterations	-n	Refinement iterations per item
--auto		Loop until critic satisfied per item
--max-iterations		Safety cap for --auto
--optimize		Input optimization per item
--format	-f	png, jpeg, or webp
--save-prompts / --no-save-prompts		Persist prompts (default: on, same as plot)
--venue		Venue style pack: built-in (neurips, icml, acl, ieee), a user pack, or custom
--aspect-ratio	-ar	Default aspect ratio when not set in the manifest
--verbose	-v	Verbose logging

paperbanana orchestrate -- Full-Paper Figure Package

Generate a publication-focused figure bundle from a full paper source, with optional data-driven plots. The command:

• parses the paper (.txt, .md, or .pdf)
• plans multiple methodology figures from section structure
• optionally discovers CSV/JSON files to plan statistical plots
• runs generation for all planned items
• writes a package folder containing figure_package.json, figures/, figures.tex, and captions.md

paperbanana orchestrate \
  --paper paper.pdf \
  --data-dir ./results \
  --max-method-figures 4 \
  --max-plot-figures 3 \
  --optimize

Use --dry-run to only plan and inspect orchestration_plan.json without API calls. Use --resume-orchestrate <id-or-path> to continue an interrupted orchestration from checkpoint state.

Flag	Description
--paper / -p	Paper source path (.txt, .md, or .pdf)
--resume-orchestrate	Resume an existing orchestration by ID or directory
--retry-failed	When resuming, include previously failed tasks
--max-retries	Extra retries per task after first failure
--data-dir	Optional directory containing CSV/JSON files for plot planning
--output-dir / -o	Parent output directory (creates orchestrate_*)
--max-method-figures	Max methodology figures to plan/generate
--max-plot-figures	Max plot figures to plan/generate
--pdf-pages	PDF-only page selection (e.g. 1-5, 2,4,6-8)
--optimize	Enable input optimization for generated items
--iterations / -n	Refinement iterations per generated item
--auto + --max-iterations	Critic-driven auto-refine mode with safety cap
--concurrency	Parallel figure generation workers
--format / -f	Output format (png, jpeg, webp)
--dry-run	Plan package only; no generation calls

paperbanana composite -- Compose Multi-Panel Figures

Stitch multiple images into a single labeled figure with (a), (b), (c) sub-panel labels:

paperbanana composite \
  panel_a.png panel_b.png panel_c.png \
  --layout 1x3 \
  --output figure2.png

Flag	Short	Description
IMAGES		Positional: paths to images to compose
--layout	-l	Grid layout: RxC (e.g. 1x3, 2x2) or auto (default: auto)
--labels		Comma-separated labels, or none to disable (default: auto (a),(b),...)
--spacing	-s	Pixel spacing between panels (default: 20)
--label-position		top or bottom (default: bottom)
--label-font-size		Font size for labels (default: 32)
--output	-o	Output path (default: composite_output.png)

This command works on any existing images — no API calls needed. It is also triggered automatically when a batch manifest includes a composite section (see paperbanana batch above).

paperbanana evaluate -- Quality Assessment

Comparative evaluation of a generated diagram against a human reference using VLM-as-a-Judge:

paperbanana evaluate \
  --generated diagram.png \
  --reference human_diagram.png \
  --context method.txt \
  --caption "Overview of our framework"

Flag	Short	Description
--generated	-g	Path to generated image (required)
--reference	-r	Path to human reference image (required)
--context		Path to source context text file or PDF (required)
--caption	-c	Figure caption (required)
--pdf-pages		PDF context only: 1-based page selection (default: all)

Scores on 4 dimensions (hierarchical aggregation per the paper):

• Primary: Faithfulness, Readability
• Secondary: Conciseness, Aesthetics

paperbanana polish -- Refine an Existing Figure

Bring your own figure: a VLM audits it against the venue style guide and proposes up to 10 concrete, actionable improvements, then an image-edit capable provider applies them to the original figure (guided edit). Suggestions are printed to the console so you can see exactly what changed.

paperbanana polish --input figure.png
paperbanana polish --input figure.png --venue icml --iterations 2 --output polished.png

Flag	Short	Description
--input	-i	Path to the existing figure image (required)
--output	-o	Output path (default: outputs/polish_<timestamp>/final_output.png)
--venue		Venue style guide: neurips (default), icml, acl, ieee, custom
--iterations	-n	Polish rounds; each round suggests and applies improvements on the previous result (default: 1)
--aspect-ratio	-ar	Target aspect ratio (default: preserve the input figure's ratio)
--num-candidates	-k	Apply each round's suggestions N times in parallel (1-8)
--budget		Budget cap in USD; polishing stops gracefully when exceeded
--seed		Random seed for reproducible edits

Requires an image provider that supports guided image edits (Google Gemini image models). If the figure already conforms to the style guide, polish exits without making changes.

paperbanana studio -- Local web UI

Requires pip install 'paperbanana[studio]' (Gradio).

paperbanana studio
paperbanana studio --port 8080 --output-dir ./my_outputs

Flag	Description
--host	Bind address (default 127.0.0.1)
--port	Port (default 7860)
--share	Create a temporary public Gradio link (do not use with sensitive data)
--config	Path to YAML config
--output-dir / -o	Default output directory for runs
--root-path	URL subpath when behind a reverse proxy

paperbanana setup -- First-Time Configuration

paperbanana setup

Interactive wizard that first asks whether to use the official Gemini API. If you choose official API, it follows the default AI Studio key flow; if not, it asks for a custom Gemini-compatible URL and API key.

paperbanana data -- Reference Dataset

# Download the PaperBananaBench reference set (~254 MB, one command)
paperbanana data download

# Import plot references too (or both)
paperbanana data download --task plot
paperbanana data download --task both

# Inspect / clear the cache
paperbanana data info
paperbanana data clear

The dataset is served from a project-hosted GitHub release mirror (bench-data-v1) and its SHA256 checksum is verified before extraction. Credit to the PaperBananaBench authors — the mirror tracks their 2026-03-22 revision. The set is cached under ~/.cache/paperbanana/ (override with PAPERBANANA_CACHE_DIR); generation commands can also fetch it on first use via --auto-download-data.

---

Python API

import asyncio
from paperbanana import PaperBananaPipeline, GenerationInput, DiagramType
from paperbanana.core.config import Settings

settings = Settings(
    vlm_provider="openai",
    vlm_model="gpt-5.2",
    image_provider="openai_imagen",
    image_model="gpt-image-1.5",
    optimize_inputs=True,   # Enable input optimization
    auto_refine=True,       # Loop until critic is satisfied
)

pipeline = PaperBananaPipeline(settings=settings)

result = asyncio.run(pipeline.generate(
    GenerationInput(
        source_context="Our framework consists of...",
        communicative_intent="Overview of the proposed method.",
        diagram_type=DiagramType.METHODOLOGY,
    )
))

print(f"Output: {result.image_path}")

Progress callbacks: generate() and continue_run() accept an optional progress_callback argument. The pipeline invokes it with PipelineProgressEvent objects (stage, message, seconds, iteration, extra) at each step (optimizer, retriever, planner, stylist, visualizer, critic), so you can show progress in UIs or log timing without patching agents.

To continue a previous run:

from paperbanana.core.resume import load_resume_state

state = load_resume_state("outputs", "run_20260218_125448_e7b876")
result = asyncio.run(pipeline.continue_run(
    resume_state=state,
    additional_iterations=3,
    user_feedback="Make the encoder block more prominent",
))

See examples/generate_diagram.py and examples/generate_plot.py for complete working examples.

---

MCP Server

PaperBanana includes an MCP server for use with Claude Code, Cursor, or any MCP-compatible client. Add the following config to use it via uvx without a local clone:

{
  "mcpServers": {
    "paperbanana": {
      "command": "uvx",
      "args": ["--from", "paperbanana[mcp]", "paperbanana-mcp"],
      "env": { "GOOGLE_API_KEY": "your-google-api-key" }
    }
  }
}

Eleven MCP tools are exposed: generate_diagram, generate_plot, continue_run (resume a prior run_* with optional feedback), continue_diagram, continue_plot, evaluate_diagram, evaluate_plot, orchestrate_figures (full-paper figure packages), batch_diagrams, batch_plots, and download_references.

The repo also ships with 3 Claude Code skills:

• /generate-diagram <file> [caption] - generate a methodology diagram from a text file
• /generate-plot <data-file> [intent] - generate a statistical plot from CSV/JSON data
• /evaluate-diagram <generated> <reference> - evaluate a diagram against a human reference

See mcp_server/README.md for full setup details (Claude Code, Cursor, local development).

---

Overleaf Integration (GitHub Action)

Keep your paper's methodology figure in sync with the text — automatically. PaperBanana ships a GitHub Action that pairs with Overleaf's built-in GitHub sync: push your .tex changes, the action extracts the methodology section, generates the figure, and commits back the image plus a ready-to-\input LaTeX snippet. Pull in Overleaf and it's in your file tree.

- uses: actions/checkout@v4
- uses: llmsresearch/paperbanana/integrations/github-action@main
  with:
    tex-file: sections/method.tex
    caption: "Overview of our proposed framework"
  env:
    OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

See integrations/github-action/README.md for the full workflow, all inputs, and cost-control options.

---

Configuration

Default settings are in configs/config.yaml. Override via CLI flags or a custom YAML:

paperbanana generate \
  --input method.txt \
  --caption "Overview" \
  --config my_config.yaml

Key settings:

vlm:
  provider: openai           # openai, atlas, gemini, or openrouter
  model: gpt-5.2

image:
  provider: openai_imagen    # openai_imagen, atlas_imagen, google_imagen, or openrouter_imagen
  model: gpt-image-1.5

pipeline:
  num_retrieval_examples: 10
  refinement_iterations: 3
  # auto_refine: true        # Loop until critic is satisfied
  # max_iterations: 30       # Safety cap for auto_refine mode
  # optimize_inputs: true    # Preprocess inputs for better generation
  output_resolution: "2k"

reference:
  path: data/reference_sets

output:
  dir: outputs
  save_iterations: true
  save_metadata: true

Environment variables (.env):

# OpenAI (default)
OPENAI_API_KEY=your-key
OPENAI_BASE_URL=https://api.openai.com/v1    # or Azure endpoint
OPENAI_VLM_MODEL=gpt-5.2                      # override model
OPENAI_IMAGE_MODEL=gpt-image-1.5              # override model

# Atlas Cloud
ATLASCLOUD_API_KEY=your-key
ATLASCLOUD_BASE_URL=https://api.atlascloud.ai/v1
ATLASCLOUD_VLM_MODEL=deepseek-ai/DeepSeek-V3-0324
ATLASCLOUD_IMAGE_BASE_URL=https://api.atlascloud.ai/api/v1
ATLASCLOUD_IMAGE_MODEL=openai/gpt-image-2/text-to-image

# Google Gemini (alternative, free)
GOOGLE_API_KEY=your-key
GOOGLE_BASE_URL=                            # optional custom Gemini-compatible endpoint
GOOGLE_VLM_MODEL=gemini-2.5-flash          # override Gemini VLM model
GOOGLE_IMAGE_MODEL=gemini-3-pro-image-preview  # override Gemini image model

---

Project Structure

paperbanana/
├── paperbanana/
│   ├── core/          # Pipeline orchestration, types, config, resume, utilities
│   ├── agents/        # Optimizer, Retriever, Planner, Stylist, Visualizer, Critic
│   ├── providers/     # VLM and image gen provider implementations
│   │   ├── vlm/       # OpenAI, Atlas Cloud, Gemini, OpenRouter VLM providers
│   │   └── image_gen/ # OpenAI, Atlas Cloud, Gemini, OpenRouter image gen providers
│   ├── reference/     # Reference set management (13 curated examples)
│   ├── guidelines/    # Style guidelines loader
│   └── evaluation/    # VLM-as-Judge evaluation system
├── configs/           # YAML configuration files
├── prompts/           # Prompt templates for all agents + evaluation
│   ├── diagram/       # context_enricher, caption_sharpener, retriever, planner, stylist, visualizer, critic
│   ├── plot/          # plot-specific prompt variants
│   └── evaluation/    # faithfulness, conciseness, readability, aesthetics
├── data/
│   ├── reference_sets/  # 13 verified methodology diagrams
│   └── guidelines/      # NeurIPS-style aesthetic guidelines
├── examples/          # Working example scripts + sample inputs
├── scripts/           # Data curation and build scripts
├── tests/             # Test suite
├── mcp_server/        # MCP server for IDE integration
└── .claude/skills/    # Claude Code skills (generate-diagram, generate-plot, evaluate-diagram)

Development

# Install with dev dependencies
pip install -e ".[dev,openai,google]"

# Run tests
pytest tests/ -v

# Lint
ruff check paperbanana/ mcp_server/ tests/ scripts/

# Format
ruff format paperbanana/ mcp_server/ tests/ scripts/

Citation

This is an unofficial implementation. If you use this work, please cite the original paper:

@article{zhu2026paperbanana,
  title={PaperBanana: Automating Academic Illustration for AI Scientists},
  author={Zhu, Dawei and Meng, Rui and Song, Yale and Wei, Xiyu
          and Li, Sujian and Pfister, Tomas and Yoon, Jinsung},
  journal={arXiv preprint arXiv:2601.23265},
  year={2026}
}

Original paper: https://arxiv.org/abs/2601.23265

Disclaimer

This project is an independent open-source reimplementation based on the publicly available paper. It is not affiliated with, endorsed by, or connected to the original authors, Google Research, or Peking University in any way. The implementation may differ from the original system described in the paper. Use at your own discretion.

License

MIT