PaperBanana-CN: An Open-Source Scientific Illustration Tool Supporting Custom API Integration

Developer Mylszd recently open-sourced PaperBanana-CN on the Linux.do community. It is a Chinese-language scientific illustration generation tool based on a modification of PaperBanana. The core improvement is support for custom model API integration, removing reliance on official Gemini or OpenAI interfaces and directly solving the proxy configuration challenges faced by Chinese users.

The project’s motivation is very pragmatic: when creating figures for papers, the author likes to run a few drafts using large models to find ideas. However, most open-source scientific figure-generation tools either hardcode the API or have proxy setup processes so complex that they discourage use. Especially after the release of GPT Image 2, many projects still haven’t caught up. PaperBanana-CN instead hands full control of API configuration to the user — if you have your own model URL and API Key, just fill them into the webpage to start using it.

Problems It Solves

The core logic of a scientific figure-generation tool is a multi-agent pipeline: input the paper’s methods section and figure captions, and the model automatically generates multiple candidate visualizations, which users can then filter and export in bulk. The original PaperBanana already implemented this, but there were two significant usability barriers:

1. Hardcoded APIs: the original version only supported Gemini or OpenAI official endpoints, forcing domestic users either to use VPNs or manually modify the code to connect through proxies.
2. High configuration threshold: even if custom APIs were supported, setup was unintuitive, requiring editing config files or environment variables.

The PaperBanana-CN improvements are straightforward: add Base URL and API Key input fields directly in the web interface, supporting any OpenAI-compatible API. This means you can use interfaces from aggregator platforms like OpenAI Hub, your own locally deployed model service, or even any third-party proxy. The API Key is stored locally for reuse.

Adapted for GPT Image 2, But Not the Optimal Choice

PaperBanana-CN fully supports OpenAI’s GPT Image 2, including features such as streaming previews, automatic Responses API fallback, reference image editing, and custom dimensions.
However, according to testing, GPT Image 2 underperforms compared to Banana Pro (another evolution of PaperBanana) in this context.

The reason lies in overlapping capabilities: GPT Image 2 excels at text fidelity and zero-shot image generation, while the PaperBanana pipeline already focuses on textual stability and stylistic consistency. Their strengths overlap, offering limited complementarity.
The developer’s recommendation: if you have access to Banana Pro, use it first; GPT Image 2 works but delivers lower-quality results.

This is a pragmatic judgment — the primary need for scientific figures is consistent style, clear labels, and adherence to academic standards, not artistic expressiveness. GPT Image 2 performs excellently in general creative tasks, but in this scientific niche, domain-tuned models remain superior.

Typical Use Cases

The developer lists several common scenarios:

• Method diagrams: input the methods section text to generate flowcharts or architecture diagrams.
• Experimental result visualization: generate charts or comparison figures based on data descriptions.
• Conceptual illustrations: convert abstract concepts into intuitive visualizations.

These all demand fast iteration. Traditionally, researchers must draw drafts themselves or work with designers, a long and costly process. Using large models to generate candidate figures compresses the iteration cycle to minutes — researchers just pick and tweak from the outputs.

Current Limitations

The developer also candidly lists known issues:

1. Model capability dependency: output quality depends entirely on the model you connect. Weaker models produce noticeably worse results.
2. Incomplete Chinese support: though the interface is localized, model understanding of Chinese academic terms still lags behind English.
3. Consistency issues: generated images can vary in style, requiring manual selection.
4. Limited complex chart support: for data-heavy or highly relational charts, outputs may be unstable.

Some issues can be resolved at the tool level (e.g., adding a “style lock” feature), while others stem from model limitations (e.g., Chinese comprehension and complex chart generation). Users are encouraged to report issues and suggestions via GitHub.

Technical Implementation Details

At its core, PaperBanana-CN uses a multi-agent pipeline, a methodology originally developed by the PaperBanana author during an internship at Google, which has since patented the relevant workflow.
Thus, the pipeline cannot be used commercially, and PaperBanana-CN, as an open-source project, is subject to the same restriction.

Architecturally, PaperBanana-CN introduces these improvements:

• API abstraction layer: unifies model call logic under the OpenAI-style interface standard.
• Local storage: API Keys and settings remain local, not uploaded.
• Chinese interface: all UI text and prompts are localized.
• Batch export: one-click packaging and download of all candidate outputs.

Installation is straightforward, supporting both pip and uv:

# Install via pip
pip install paperbanana-cn

# Or install via uv
uv tool install paperbanana-cn

After installation, start it from the command line; a web interface will open automatically. Enter your Base URL and API Key on first use — they’ll be saved thereafter.

Relationship to PaperBanana-Pro

The PaperBanana ecosystem has two main evolutionary branches: PaperBanana-Pro and PaperBanana-CN, each serving different purposes:

• PaperBanana-Pro: developed by elpsykongloo, refined through 21 engineering iterations and 70+ unit tests. It evolved from an academic prototype into a production-ready tool, featuring background generation, a drawing workspace, portable result packages, and more advanced features.
• PaperBanana-CN: developed by Mylszd, focuses on simplifying usage for Chinese users and proxy setups. It’s lighter but more flexible and user-friendly.

If you need a complete workflow for scientific illustration, PaperBanana-Pro is the better choice.
If you just want to quickly generate a few candidate figures or integrate your own model service, PaperBanana-CN is more suitable.

Insights for Developers

PaperBanana-CN offers a valuable design lesson: many open-source tools default to binding official APIs — reasonable for global users, but a barrier for Chinese users. Handing API configuration power to users both lowers usability hurdles and expands applicability.

The cost of such modification is low. The core idea is to create a solid API abstraction layer, decoupling model calls from business logic. As long as the underlying model adheres to the OpenAI format (now the de facto standard), integration is seamless.
For scenarios needing multiple model backends, aggregator platforms like OpenAI Hub can further simplify setup — one key for all models, no separate configuration needed.

Another aspect worth attention is the open-source license. PaperBanana’s core methodology is covered by Google’s patent, prohibiting commercial use. This isn’t a concern for individual researchers or academic institutions, but commercialization would require either a new design or official authorization.

Summary

PaperBanana-CN addresses a very specific pain point: enabling Chinese users to easily generate scientific illustrations using their own model services. The improvements are simple yet practical.
If you frequently need figures for papers or want to quickly validate visual concepts, this tool is worth a try.

The project is open-sourced on GitHub, featuring good code quality and documentation. The developer is active in the community and responsive to issues.
The only caution is the commercial use restriction — ensure licensing compliance if your application is commercial.

---

References

• PaperBanana-CN GitHub Repository – Source code and documentation
• Linux.do Discussion Thread – Developer’s project announcement and usage guide
• PaperBanana-Pro GitHub Repository – Another evolution with a more complete feature set