PaperBanana: The AI That's Automating Academic Illustration (And It's Kind of Mind-Blowing)

If you've ever written a research paper, you know the pain: you've done the hard work, written thousands of words explaining your groundbreaking methodology, and then... you need to create diagrams. Beautiful, publication-ready diagrams that somehow capture your complex ideas in a single visual. For many researchers, this becomes the most time-consuming part of the entire process.

Enter PaperBanana, a revolutionary framework from researchers at Peking University and Google Cloud AI Research that's tackling this exact bottleneck. And yes, they named it PaperBanana because even serious AI research deserves a smile.

What Makes PaperBanana Special?

Think of PaperBanana as your personal illustration team, but instead of humans, it's five specialized AI agents working together. Each agent has a specific role: the Retriever finds relevant reference examples from existing papers, the Planner translates your research context into detailed visual descriptions, the Stylist ensures everything looks professionally polished, the Visualizer creates the actual diagrams, and the Critic reviews and refines the output until it meets publication standards.

This isn't just about slapping together some boxes and arrows. PaperBanana generates diagrams that are faithful to your research, concise enough to be readable, aesthetically pleasing, and sophisticated enough to appear in top-tier conferences like NeurIPS.

PaperBanana's architecture: Five specialized AI agents collaborate to transform research content into publication-ready illustrations.

The Secret Sauce: Reference-Driven Intelligence

What sets PaperBanana apart is its reference-driven approach. Instead of generating illustrations from scratch with no context, it learns from the visual language already established in academic publishing. The system analyzes methodology diagrams from recent NeurIPS papers, understanding not just what makes a diagram functional, but what makes it beautiful and publication-ready.

The results speak for themselves. In comprehensive testing against leading baselines, PaperBanana consistently outperformed competitors across all evaluation dimensions: faithfulness, conciseness, readability, and aesthetics. It's not just good—it's setting a new standard.

Beyond Methodology Diagrams

But here's where it gets even more interesting: PaperBanana doesn't just do methodology diagrams. It also generates high-quality statistical plots. The researchers tested both code-based and image generation approaches for creating visualizations, revealing fascinating trade-offs. Image generation creates more visually appealing plots, but code-based methods maintain better content fidelity. Understanding these nuances helps researchers choose the right approach for their needs.

The Benchmark That Changes Everything

To properly evaluate automated illustration generation, the team created PaperBananaBench—a rigorous benchmark comprising 292 test cases curated from NeurIPS 2025 publications. This benchmark captures the sophisticated aesthetics and diverse logical compositions of modern AI research, spanning multiple research domains and illustration styles.

The average source context contains over 3,000 words, proving that PaperBanana can handle the complexity of real research papers, not just simplified examples.

PaperBananaBench statistics showing 292 test cases with average source context of 3,020 words per diagram.

PaperBanana consistently outperforms baselines across all evaluation dimensions: faithfulness, conciseness, readability, and aesthetics.

Real-World Applications

The practical applications extend beyond just generating new diagrams. PaperBanana can enhance the aesthetics of existing human-drawn diagrams, applying automatically summarized style guidelines to elevate visual quality. Imagine taking a rough sketch and having it instantly transformed into a polished, publication-ready illustration that maintains your original intent while looking professionally designed.

Before and after: PaperBanana transforms verbose, outdated diagrams into concise, aesthetically modern illustrations while maintaining accuracy.

The Road Ahead

Of course, no system is perfect. The researchers openly acknowledge failure modes, particularly around connection errors in complex diagrams. But this transparency is refreshing—they're not claiming to have solved everything, just to have made a significant leap forward.

For AI researchers, content creators, and anyone involved in scientific communication, PaperBanana represents something bigger than just a tool. It's a glimpse into a future where the tedious parts of research communication are automated, freeing scientists to focus on what they do best: pushing the boundaries of knowledge.

The code is available on GitHub, the paper is on arXiv, and the framework is ready to explore. As AI continues to augment scientific workflows, tools like PaperBanana remind us that automation isn't about replacing human creativity—it's about amplifying it, one beautifully generated diagram at a time.

GPT-4o aces its multimodal classmates—but still can’t dethrone specialist vision models

 OpenAI’s GPT-4o may be the first flagship model to unify text, image and audio in a single stack, but a new EPFL benchmarking effort shows just how far even the best “everything” model still lags behind purpose-built computer-vision networks. In “How Well Does GPT-4o Understand Vision? Evaluating Multimodal Foundation Models on Standard Computer Vision Tasks,” researchers tested GPT-4o alongside six other foundation models—o4-mini, Gemini 1.5 Pro, Gemini 2.0 Flash, Claude 3.5 Sonnet, Qwen2-VL and Llama 3.2—on six bread-and-butter CV tasks that every undergrad knows: ImageNet-style classification, MS-COCO object detection, semantic segmentation, instance grouping, monocular depth and surface-normal prediction.

Turning text-only giants into pixel workers

Most API-level models can’t output polygons or depth maps, so the team invented a prompt-chaining framework that decomposes each vision problem into a sequence of classification subtasks that any chatty LLM can answer. A recursive “zoom-and-vote” routine localises objects, SLIC superpixels stand in for pixels in segmentation, and pairwise ranking lets the models infer relative depth.

Key takeaways

Finding	What happened	Why it matters
Generalist, not specialist	All MFMs landed well below state-of-the-art CV models on every benchmark.	Even massive cross-modal pre-training doesn’t yet replace task-specific supervision.
Semantic > geometric	Scores on classification, detection and segmentation were much higher than on depth or normals.	MFMs learn semantics from caption data but have little innate 3-D understanding.
GPT-4o still best of breed	GPT-4o topped the non-reasoning field in four of six tasks.	Its larger context window and image-generation head translate into better pixel comprehension.
Reasoning helps with 3-D	Smaller “o3” reasoning models edged GPT-4o on depth and normals.	Structured chain-of-thought may compensate for weaker raw vision priors.
Prompt sensitivity drops with quality	Higher-capacity models varied less when the researchers tweaked prompt chains.	Robustness could become a practical proxy for measuring model quality without labels.

The bigger picture

For product builders eyeing GPT-4o as a drop-in object detector, the study is a sobering reality check; you’ll still need a Mask R-CNN or SAM in the loop for pixel-perfect jobs. But the results also highlight the upside of super-general models: with zero fine-tuning and only clever prompting, GPT-4o can solve half a dozen vision tasks “well enough”—a compelling baseline for multimodal agents that prefer breadth over razor-edge accuracy.

The authors have open-sourced their fm-vision-evals framework so future models can be dropped into the same gauntlet—no weight access required. Expect the next wave of Gemini, Claude and Llama releases to cite these scores the way language-model papers brag about MMLU.

Paper link: arXiv 2507.01955 (PDF)