Academic Figure Generation

Academic Figure Generation

Thin CLI wrapper around PaperBanana (a.k.a. PaperVizAgent), a multi-agent figure-generation pipeline for academic papers.

The skill provides exactly one script: scripts/generate.py. It feeds your method text + caption into PaperBanana and writes N candidate PNGs. Model selection and API keys come from PaperBanana's own configs/model_config.yaml — the wrapper does not override them.

One-time setup

1. Clone PaperBanana somewhere convenient:

   bashCopy code
   git clone https://github.com/dwzhu-pku/PaperBanana.git ~/PaperBanana
   cd ~/PaperBanana
   uv venv && uv pip install -r requirements.txt
   

2. Configure configs/model_config.yaml — set the image model and the matching API key. Two common setups:

   yamlCopy code
   defaults:
     image_model_name: "gemini-3-pro-image-preview"   # or "openai/gpt-5.4-image-2"
     model_name: "gemini-3.1-pro-preview"             # text model for Planner/Stylist/Critic
   
   api_keys:
     google_api_key: "..."        # required for Gemini models
     openrouter_api_key: ""       # required for openai/gpt-5.4-image-2
   

   Use Gemini if you have a Google AI key; use GPT-Image-2 via OpenRouter if you have an OpenRouter key. Pick one — there's nothing else to wire up.

Workflow

Step 1: Gather inputs

You need:

1. Method text: the relevant section of the paper describing the approach (./method.md or ./method.tex).
2. Figure caption: the target caption, e.g. "Figure 1: Overview of our framework".

If the user only gives a vague request, ask:

• What aspect of the method should the figure focus on?
• Style? (block diagram, flowchart, pipeline, architecture, comparison)
• Venue / column width? (ACL ≤ 7.5", NeurIPS single-column 5.5")

Step 2: Generate

bashCopy code
~/PaperBanana/.venv/bin/python scripts/generate.py \
  --paperbanana-root ~/PaperBanana \
  --method-file ./method.md \
  --caption "Figure 1: Overview of our framework" \
  --out-dir ./figures/v1 \
  --candidates 3 \
  --aspect-ratio 16:9

Step 3: Present & iterate

• Show all candidates to the user.
• Common refinements: color scheme, layout, label text, font size.
• Re-run with a tweaked caption or more candidates.

Step 4: Export

• PNGs are written as candidate_0.png, candidate_1.png, … in --out-dir.
• For camera-ready PDFs: magick candidate_0.png candidate_0.pdf.

Style guidelines

• Color: consistent, colorblind-friendly palette
• Fonts: match the paper's body font (Times for ACL/EMNLP, Helvetica/Arial for many ML venues)
• Labels: concise; no full sentences inside the diagram
• Arrows: solid for data flow, dashed for optional / feedback loops
• Whitespace: don't overcrowd — reviewers skim figures in seconds

Common figure types

Troubleshooting

• 429 RESOURCE_EXHAUSTED on Gemini: monthly Google AI Studio spending cap hit. Raise it at https://ai.studio/spend or switch image_model_name to openai/gpt-5.4-image-2 and set OPENROUTER_API_KEY.
• OpenRouter Client not initialized: OPENROUTER_API_KEY not in env and openrouter_api_key not in yaml.
• No PNGs in output dir: check out_dir/results.json for the raw per-candidate response and any error messages.
• Long latency (>5 min): most wall time is the image model. Lower --candidates or use --exp-mode vanilla for faster iteration.

Links

• PaperBanana repo: https://github.com/dwzhu-pku/PaperBanana
• PaperVizAgent (Google Research version of the same project): https://github.com/google-research/papervizagent