PaperBanana: Automating Academic Illustration for AI Scientists (huggingface.co)

The introduction of PaperBanana marks a significant advancement in the automation of academic illustration, specifically for AI researchers. This agentic framework utilizes advanced vision-language models (VLMs) and image generation techniques to eradicate the time-consuming bottleneck of creating publication-ready illustrations. PaperBanana orchestrates a series of specialized agents that can autonomously retrieve references, plan the visual content and style, render images, and refine them through self-critique. This innovation is validated using PaperBananaBench, a comprehensive benchmarking suite consisting of 292 test cases derived from NeurIPS 2025 publications, showcasing its superior performance in terms of faithfulness, conciseness, readability, and aesthetics. For the AI/ML community, PaperBanana encapsulates a transformative approach, paving the way for more efficient research workflows and enabling scientists to focus on content rather than the intricacies of visual representation. Its implications extend beyond academia, emphasizing its potential applications in technical blogging and other fields requiring structured visual data. The ability to automate high-quality statistical plots and diverse illustration styles enhances the overall dissemination of scientific knowledge, ultimately contributing to a more connected and informed research ecosystem.