PaperBanana Skills

<p align="center"> <a href="https://github.com/PlutoLei/paperbanana-skill/stargazers"><img alt="GitHub Stars" src="https://img.shields.io/github/stars/PlutoLei/paperbanana-skill?style=flat-square&color=yellow" /></a> <img alt="Version" src="https://img.shields.io/badge/version-4.4.0-blue?style=flat-square" /> <img alt="Agent Skills" src="https://img.shields.io/badge/Agent%20Skills-Standard-2B6CB0?style=flat-square" /> <img alt="Multi-Runtime" src="https://img.shields.io/badge/Runtime-Multi-success?style=flat-square" /> <img alt="Python" src="https://img.shields.io/badge/Python-3.10+-3776AB?style=flat-square&logo=python&logoColor=white" /> <img alt="Providers" src="https://img.shields.io/badge/Providers-8-green?style=flat-square" /> <img alt="GPT Image 2" src="https://img.shields.io/badge/GPT%20Image%202-native-blueviolet?style=flat-square" /> <img alt="Eval" src="https://img.shields.io/badge/Eval-6--item%20Checklist-orange?style=flat-square" /> <a href="https://github.com/PlutoLei/paperbanana-skill/blob/master/LICENSE"><img alt="MIT License" src="https://img.shields.io/badge/License-MIT-black?style=flat-square" /></a> </p> <p align="center"> <strong>One sentence in, publication-quality academic figure out.</strong><br/> Powered by a 5-agent pipeline that plans, styles, generates, and self-critiques your illustrations. </p> <p align="center"> <strong>English</strong> | <a href="README_CN.md">中文</a> </p>

---

Gallery

<table> <tr> <td align="center"><strong>Biology — Signal Pathway</strong><br/><img src="examples/bio_signaling_pathway.png" width="400"/></td> <td align="center"><strong>NLP — RAG Pipeline</strong><br/><img src="examples/nlp_rag_pipeline.png" width="400"/></td> </tr> <tr> <td align="center"><strong>Data Engineering — Lakehouse</strong><br/><img src="examples/data_lakehouse.png" width="400"/></td> <td align="center"><strong>Medical AI — U-Net + Mamba</strong><br/><img src="examples/unet_mamba_segmentation.png" width="400"/></td> </tr> <tr> <td align="center"><strong>Medical Imaging — TextMamba3D</strong><br/><img src="examples/textmamba3d_architecture.png" width="400"/><br/>_{<em>gpt-image-2 · paper-grade info density</em>}</td> <td align="center"><strong>Game Theory — Influence Diagram</strong><br/><img src="examples/game_theory_influence_diagram.png" width="400"/><br/>_{<em>Gemini · soft pastel academic aesthetic</em>}</td> </tr> <tr> <td align="center"><strong>Ablation Study — BraTS 2023</strong><br/><img src="examples/ablation_brats_figure.png" width="400"/><br/>_{<em>gpt-image-2 · 2×2 MRI panels + Dice bar chart</em>}</td> <td align="center"><strong>Scientific Slide — scRNA-seq Workflow</strong><br/><img src="examples/slide_scrna_workflow.png" width="400"/><br/>_{<em>paperbanana-slide-deck · single-cell analysis pipeline</em>}</td> </tr> </table> <p align="center"><em>All figures generated from plain text descriptions — zero manual drawing.</em></p>

Slide Deck Showcase — "The Flywheel Learning Method"

A real 10-slide lecture deck built with paperbanana-slide-deck. Below: 4 selected slides showing set-wide style consistency (same warm off-white palette, sketch-notes hand-drawn typography, and gear motif across the whole deck).

<table> <tr> <td align="center"><strong>Slide 1 — Cover</strong><br/><img src="examples/slide_flywheel_01_cover.png" width="380"/></td> <td align="center"><strong>Slide 4 — Flywheel Model</strong><br/><img src="examples/slide_flywheel_04_model.png" width="380"/></td> </tr> <tr> <td align="center"><strong>Slide 7 — AI Tools Do's and Don'ts</strong><br/><img src="examples/slide_flywheel_07_ai_tools.png" width="380"/></td> <td align="center"><strong>Slide 10 — Let the Flywheel Spin</strong><br/><img src="examples/slide_flywheel_10_summary.png" width="380"/></td> </tr> </table> <p align="center"><em>One command: <code>paperbanana-slide-deck</code> picks a style preset, plans the outline, drafts per-slide prompts, then generates all slides with consistent design tokens.</em></p> <p align="center">_{The same pipeline now routes across <strong>8 providers</strong> — render this deck on <code>gpt-image-2</code> (clean Chinese titles), <code>gemini</code> (fast &amp; cheap), or any of <strong>100+ LiteLLM backends</strong> / local <code>ollama</code> models, with no workflow change.}</p> <details> <summary><strong>More Examples</strong> (architecture diagrams, traditional aesthetics)</summary> <br/> <table> <tr> <td align="center"><strong>Transformer Architecture</strong><br/><img src="examples/transformer_architecture.png" width="380"/></td> <td align="center"><strong>Mamba SSM Architecture</strong><br/><img src="examples/mamba_architecture.png" width="380"/></td> </tr> <tr> <td align="center" colspan="2"><strong>RAG Pipeline</strong><br/><img src="examples/rag_pipeline.png" width="760"/></td> </tr> <tr> <td align="center" colspan="2"><strong>Chinese Calligraphy — 自律 (Self-Discipline)</strong><br/><img src="examples/chinese_calligraphy.png" width="560"/><br/>_{<em>Gemini · bold expressive brushwork + 飞白 on xuan paper</em>}</td> </tr> </table> </details>

---

Skills in this Marketplace

| Skill | Scope | Description | Version | |-------|-------|-------------|---------| | paperbanana | user | Academic diagrams, plots, slides, and quality evaluation | v4.0.0 | | paperbanana-slide-deck | project | Full slide deck orchestration (RDIV workflow) + 150+ style presets | v1.1.0 |

Feature Matrix

| Capability | Status | Details | |------------|--------|---------| | GPT Image 2 native support | ✅ v4.3 New | gpt-image-2 (2026-04-21) with true 16:9 up to 2048×1152, quality tier (low/medium/high), full RDIV pipeline + Critic | | Smart provider routing | ✅ v4.3 New | Auto-pick openai vs gemini by scenario; explicit 用 GPT/用 Gemini/两路并行 override always respected | | Methodology diagrams | ✅ | Text → publication-quality figure in 30s | | Statistical plots | ✅ | CSV/JSON data → auto-styled academic plot | | Presentation slides | ✅ | Markdown → 4K slide with 150+ style presets | | Multi-venue styles | ✅ New | --venue neurips\|icml\|acl\|ieee\|custom | | PDF input | ✅ New | --input paper.pdf --pages 3-5 | | 6-item quality eval | ✅ New | Binary checklist: completeness, layout, annotation, color, legibility, hallucination | | Autoresearch loop | ✅ New | Automated prompt self-optimization with keep/revert | | Error handling | ✅ New | Critic UNREVIEWED status, provider fallback chains, retry filtering | | 8 VLM providers | ✅ | Gemini, Claude, OpenAI, Bedrock, OpenRouter + LiteLLM (100+ backends), Ollama (local models), claude_code (via claude CLI) | | Auto-refine | ✅ | --auto loops until Critic is satisfied | | Run continuation | ✅ | --continue with --feedback for iterative refinement | | Dynamic aspect ratio | ✅ | 8 Imagen ratios, Planner auto-recommends |

---

What's New in v4.4 — Wave-Parallel Slide Batches

slide-batch now generates slides concurrently: each slide gets its own pipeline instance (isolated run directory, isolated Critic loop), with a 5s start-up stagger, in-batch delayed retry for transient 503s, and an end-of-batch serial retry pass for stragglers.

Measured: 6 slides at --concurrent 3 in 309s vs a 768s serial estimate — 0.40x wall-clock (~2.5x speedup), zero dropped slides, identical per-slide quality gating.

Also in 4.4:

• Smarter delivery — the final image per slide is the highest-critic-score iteration rather than simply the last one; critic_score_threshold=9.0 skips provably-done rounds early (calibrated on 69 historical runs with zero false early-stops).
• Auto-routing decision table and the X_imagen provider-naming warning are now part of SKILL.md.

The runtime features (--concurrent, argmax delivery, threshold early-stop, slide generation itself) live in the maintained fork PlutoLei/paperbanana — see Quick Start. On upstream builds, slide-batch runs serially.

---

What's New in v4.3 — GPT Image 2 First-Class Support

OpenAI released gpt-image-2 on 2026-04-21. PaperBanana v4.3 integrates it natively so the full Retriever → Planner → Stylist → Visualizer → Critic pipeline runs on gpt-image-2 outputs. You get quality-gated images at up to 2048×1152 without leaving paperbanana.

Adapter upgrade

| Feature | Before (v4.2) | After (v4.3) | |---------|---------------|--------------| | Default OpenAI model | gpt-image-1.5 | gpt-image-1.5 — but gpt-image-2 is now fully wired in too | | Output sizes | 1024×1024 / 1536×1024 / 1024×1536 (3 sizes) | Adds 2048×1152 (true 16:9), 1536×1536, 1792×1024, 1152×2048 | | quality=low\|medium\|high | ❌ rejected | ✅ auto-sent for gpt-image-2 | | Supported ratios | 3 (1:1, 3:2, 2:3) | 8 (all paperbanana ratios; no more downgrade) | | Critic loop | Only on Gemini | ✅ Runs on gpt-image-2 too — catches Chinese typo bugs, missing nodes |

Switching is a two-flag change:

python -m paperbanana.cli generate \
  --image-provider openai_imagen --image-model gpt-image-2 \
  --aspect-ratio 16:9 \
  --input prompt.txt --caption "..."

Auto routing by scenario

The skill picks the right provider based on your request's signal:

| Scenario | Auto-routes to | Why | |----------|----------------|-----| | User says 用 GPT / 用 Gemini / 两路并行 | That provider (or both) | Explicit intent always wins | | --purpose submission / "投稿用" | gpt-image-2 high | Rigor priority | | Slide deck with Chinese titles | gpt-image-2 | Avoid Gemini's duplicate-character bug (see below) | | Edit with ≥ 2 reference images | gpt-image-2 | Avoid Gemini's multi-image hallucination | | Prompt mentions 山水 / 书法 / 古风 / 水墨 | gemini | Gemini dominates traditional East-Asian aesthetics | | generate with architecture / multi-stage / ablation keywords | gpt-image-2 high | GPT wins on dense multi-module figures | | Everything else | gemini medium (default) | Faster, cheaper, prettier for general work |

Routing is calibrated from a 16-prompt controlled comparison (details: docs/superpowers/specs/2026-04-23-image-router-design.md in the companion repo).

Before / After — routing in action

These pairs come from the same prompt sent to both providers. The routing table exists because each model has specific strengths and specific bugs.

1. Chinese slide titles — GPT wins (Gemini has a duplicate-character bug)

<table> <tr> <td align="center" width="50%"><strong>Gemini</strong><br/><img src="examples/routing-comparison/D2_gemini.png" width="440"/><br/><em>Title reads "飞轮模飞轮模型" — the prefix "飞轮模" is duplicated. Not viable for slide decks.</em></td> <td align="center" width="50%"><strong>gpt-image-2</strong><br/><img src="examples/routing-comparison/D2_openai.png" width="440"/><br/><em>Title renders cleanly: "飞轮模型 — 核心概念". Routing sends Chinese slides here.</em></td> </tr> </table>

2. Semantic correctness (diffusion process) — GPT wins

<table> <tr> <td align="center" width="50%"><strong>Gemini</strong><br/><img src="examples/routing-comparison/A2_gemini.png" width="440"/><br/><em>Cat images at x_0 through x_4 look identical; only x_T is noise. Semantics and visuals don't match.</em></td> <td align="center" width="50%"><strong>gpt-image-2</strong><br/><img src="examples/routing-comparison/A2_openai.png" width="440"/><br/><em>Cat actually degrades step-by-step — visually faithful to the diffusion process.</em></td> </tr> </table>

3. Traditional Chinese calligraphy — Gemini wins (bolder brushwork)

<table> <tr> <td align="center" width="50%"><strong>Gemini</strong><br/><img src="examples/routing-comparison/G2_gemini.png" width="440"/><br/><em>Bold expressive strokes with visible 飞白 and xuan-paper fiber — the prompt asked for "bold" and got it.</em></td> <td align="center" width="