paperbanana-lite

name

paperbanana-lite

description

Generate publication-quality academic methodology diagrams and statistical plots from text descriptions using a multi-agent Gemini pipeline. Use when the user wants to create scientific figures, methodology diagrams, architecture diagrams, or statistical plots for academic papers. Triggers on requests like "generate a diagram", "create a methodology figure", "make a plot for my paper", or "illustrate this architecture".

PaperBanana Lite

Single-file multi-agent pipeline for academic illustration generation using Google Gemini. Distills the full PaperBanana framework into one script.

Quick Start

# Download reference dataset (one-time setup, or done automatically on first run)
uv run scripts/paperbanana_lite.py setup

# Diagram generation
uv run scripts/paperbanana_lite.py generate \
  --input methodology.txt \
  --caption "Overview of our encoder-decoder architecture"

# Plot generation
uv run scripts/paperbanana_lite.py plot \
  --data results.json \
  --intent "Bar chart comparing model accuracy"

Environment

Requires GEMINI_API_KEY (or GOOGLE_API_KEY as fallback):

export GEMINI_API_KEY="your-key"

Install dependencies:

pip install google-genai pillow tenacity

Pipeline Overview

Phase 1 — Linear Planning:

1. Retriever — selects relevant reference examples from curated set via VLM
2. Planner — generates detailed textual description via in-context learning
3. Stylist — refines description with NeurIPS-style aesthetic guidelines

Phase 2 — Iterative Refinement (up to N iterations): 4. Visualizer — renders image (Gemini image gen for diagrams, matplotlib code for plots) 5. Critic — evaluates quality, provides revision feedback; loops back to step 4

CLI Options

setup (download references)

Downloads the curated reference dataset (~937 KB) from GitHub to a local cache. This is optional — references are fetched automatically on first generate or plot run if not already present.

Option	Required	Default	Description
--target-dir	No	~/.paperbanana/reference_sets	Where to store references

Reference lookup order during generation:

1. --reference-dir if explicitly provided
2. data/reference_sets (local repo directory)
3. ~/.paperbanana/reference_sets (shared cache)
4. Auto-downloads from GitHub if none found

generate (methodology diagrams)

Option	Required	Default	Description
--input	Yes	—	Text file with methodology section
--caption	Yes	—	Figure caption
--reference-dir	No	data/reference_sets	Reference set directory
--iterations	No	3	Refinement iterations
--output-dir	No	outputs	Output directory

plot (statistical plots)

Option	Required	Default	Description
--data	Yes	—	JSON file with raw data
--intent	Yes	—	Visual intent caption
--reference-dir	No	data/reference_sets	Reference set directory
--iterations	No	3	Refinement iterations
--output-dir	No	outputs	Output directory

Output

Results saved to outputs/run_<timestamp>/:

• final_output.png — final generated image
• planning.json — retrieved examples and description
• iter_N.png — image from each iteration
• iter_N_details.json — description and critic feedback per iteration

Reference Sets

References are automatically downloaded from GitHub on first use. You can also manually set up references:

# Explicit setup to default cache (~/.paperbanana/reference_sets)
uv run scripts/paperbanana_lite.py setup

# Setup to a custom location
uv run scripts/paperbanana_lite.py setup --target-dir ./my-references

For custom reference sets, the --reference-dir must point to a directory containing an index.json with curated examples. Each example needs: id, source_context, caption, image_path. See references/reference-set-format.md for the schema.