Author's Note: A detailed look at how PaperBanana automatically extracts aesthetic norms from 292 NeurIPS 2025 papers to polish hand-drawn sketches and rough charts into publication-grade academic illustrations, covering four major optimization dimensions: layout, color scheme, typography, and iconography.
Are your research figures suffering from these issues: crowded layouts, jarring colors, inconsistent fonts, or a chaotic mix of icon styles? These seemingly minor visual flaws are often the direct reason reviewers give the "figures need improvement" feedback. Manually adjusting PowerPoint or LaTeX charts one by one is time-consuming and makes it hard to maintain consistency. Now, the PaperBanana framework offers a brand-new solution—systematically refining existing charts based on aesthetic norms automatically extracted from 292 NeurIPS top-tier conference papers.
Core Value: By the end of this article, you'll master the complete PaperBanana figure refinement workflow—from submitting a rough sketch to obtaining publication-grade illustrations. You'll understand how the Stylist agent automatically summarizes top-conference aesthetic standards and how to achieve professional figure optimization across four key dimensions at a low cost using the Nano Banana Pro API.
Extracting Aesthetic Standards from 292 NeurIPS Papers
PaperBanana's Stylist agent systematically analyzed academic illustrations in 292 NeurIPS 2025 papers to automatically summarize a set of Aesthetic Guidelines. These guidelines cover the following core dimensions:
| Dimension | Common Standards in Top-Tier Papers | Common Sketch Issues |
|---|---|---|
| Layout | Uniform module spacing, clear hierarchy, logical information flow | Crowded or overlapping elements, messy hierarchy |
| Color Palette | Soft Tech Pastels, moderate contrast | Harsh primary colors, random palettes |
| Typography | Unified font family, distinct hierarchy, 9-11pt body text | Mixed fonts, inconsistent sizes |
| Iconography | Flat, minimalist icons, unified style | Rough hand-drawn look, inconsistent style |
| Connectors | Unified arrow styles, consistent line thickness | Varying line thickness, random arrows |
🎯 Technical Tip: If you want to quickly experience PaperBanana's diagram refinement, we recommend calling the Nano Banana Pro API via the APIYI (apiyi.com) platform. This platform provides a unified image generation interface, costing only $0.05 per call—an 80% discount off the official price ($0.234). It's perfect for batch refining multiple diagrams.
The Aesthetic Analysis Workflow of the Stylist Agent
The Stylist agent is the core component among PaperBanana's five major agents responsible for aesthetic standards. Its workflow is as follows:
Step 1: Reference Paper Analysis
The Retriever agent searches Google Scholar for top-tier papers related to the user's research direction and extracts their illustrations as aesthetic reference samples.
Step 2: Aesthetic Guideline Generation
The Stylist agent systematically analyzes these reference illustrations and automatically generates a structured Aesthetic Guideline Document, which includes:
Aesthetic Guidelines Summary:
├── Color Palette: Soft Tech Pastels (#E8F4FD, #B8E6CF, #FFE4B5...)
├── Font: Sans-serif family, Title 14pt bold, Body 10pt regular
├── Layout: Left-to-right flow, 20px module spacing
├── Icons: Flat design
## Nano Banana Pro Chart Polishing in Action: The 3-Round Iterative Refinement Process
PaperBanana's chart polishing isn't a one-and-done deal. Instead, it uses a **3-round iterative refinement cycle** powered by Visualizer + Critic agents to steadily boost quality. This mechanism mimics the "Draw → Review → Revise" workflow of a human designer.
### The Iterative Refinement Workflow
**Round 1: Initial Polishing**
The Visualizer agent performs the first round of polishing based on the aesthetic guidance document provided by the Stylist. This round mainly tackles the most obvious issues: layout adjustments, color palette replacement, and font unification.
**Round 2: Detail Optimization**
The Critic agent evaluates the Round 1 results across four dimensions (Faithfulness, Conciseness, Readability, and Aesthetics) and provides specific improvement suggestions. The Visualizer then conducts a second round of optimization, focusing on connector styles, icon consistency, and spacing tweaks.
**Round 3: Final Refinement**
The Critic evaluates the work once more, looking for any tiny details that might have been missed. The Visualizer performs the final polish to ensure every element strictly adheres to the aesthetic guidelines.
### The Critic Agent's 4-Dimensional Evaluation System
The Critic agent provides a quantitative assessment of each polishing round, with dimensions drawn directly from the PaperBananaBench benchmark:
| Evaluation Dimension | Weight | What's Evaluated | PaperBanana Improvement |
|----------|------|----------|---------------------|
| **Faithfulness** | 30% | Whether all info and scientific meaning from the original are preserved | Baseline Level |
| **Conciseness** | 25% | Whether redundant elements are removed to highlight core info | +37.2% |
| **Readability** | 25% | Text clarity and whether the information flow is intuitive | +12.9% |
| **Aesthetics** | 20% | Whether colors, layout, and fonts meet top-tier conference standards | +6.6% |
> 🚀 **Quick Start**: We recommend using the **APIYI (apiyi.com)** platform to quickly try out the chart polishing feature. Their Nano Banana Pro API works right out of the box with no complex setup required—you can have your first chart polished in under 5 minutes. Online demo tool: Image.apiyi.com
### Before & After: Polishing Results
Let's look at a typical methodology architecture diagram to see how it evolves through the 3-round iteration:
**Original Sketch**: A hand-drawn PowerPoint chart with a crowded layout, inconsistent colors, mixed fonts, and varying line thicknesses.
**After Round 1**: The layout is rearranged into a left-to-right flow, colors are replaced with "Soft Tech Pastels," and fonts are unified to Sans-serif. Visual quality improves by about 60%.
**After Round 2**: Connectors are unified into rounded arrows, icons are replaced with flat designs, and module spacing is adjusted to a uniform 20px. Visual quality improves by about 85%.
**After Round 3**: Fine-tuning text contrast, optimizing local white space, and ensuring all elements are perfectly aligned. The result reaches publication-grade quality, with a total visual improvement of about 95%.
---
## Nano Banana Pro Chart Polishing API Guide
Calling PaperBanana's chart polishing feature via the Nano Banana Pro API is simple—you just need to construct the right prompt. Here's how to do it.
### Minimal Code Example (Get started in 10 lines)
```python
import openai
client = openai.OpenAI(
api_key="YOUR_API_KEY",
base_url="https://api.apiyi.com/v1" # Using APIYI unified interface
)
response = client.images.generate(
model="nano-banana-pro",
prompt="Polish this diagram: [Describe your original chart content]. Apply NeurIPS-style aesthetic guidelines: soft tech pastel colors, sans-serif fonts, flat icons, left-to-right flow, uniform spacing 20px, clean minimalist academic style.",
n=1,
size="1024x1024"
)
print(response.data[0].url)
Chart Polishing Prompt Template
An effective polishing prompt needs to include three core elements:
Element 1: Original Chart Description
Polish this methodology diagram showing: [Encoder-Decoder architecture
with attention mechanism, 3 input nodes, 2 hidden layers, 1 output node]
Element 2: Aesthetic Guidelines
Apply aesthetic guidelines:
- Color: Soft Tech Pastels (#E8F4FD, #B8E6CF, #FFE4B5)
- Font: Sans-serif, title 14pt bold, body 10pt
- Layout: Left-to-right, 20px spacing
- Icons: Flat design, 1.5px stroke
Element 3: Preservation Constraints
Preserve: All original nodes, connections, labels, and scientific meaning.
Do NOT add or remove any information elements.
5 Common Chart Polishing Prompt Examples
| Chart Type | Prompt Keywords | Focus Area |
|---|---|---|
| Methodology Architecture | "methodology architecture diagram" | Layout + Icons |
| Algorithm Flowchart | "algorithm flowchart" | Connectors + Colors |
| Network Structure | "neural network structure" | Hierarchy + Spacing |
| System Pipeline | "system pipeline diagram" | Info Flow + Consistency |
| Encoder-Decoder Framework | "encoder-decoder framework" | Symmetry + Fonts |
💰 Cost Optimization: For research teams needing to polish charts in bulk, calling the Nano Banana Pro API via the APIYI (apiyi.com) platform can significantly cut costs. Each call is only $0.05 (about 80% off the official price of $0.234). Polishing 10 charts costs just $0.50, making it perfect for batch optimization before submitting a paper.
6 Best Practices for Nano Banana Pro Diagram Polishing
Based on PaperBanana's technical features and real-world experience, here are 6 best practices for polishing your diagrams:
Practice 1: Give as much detail as possible in your original description
The quality of the polish depends directly on how accurately your prompt describes the original diagram. We recommend including:
- All node names and hierarchical relationships
- Connection directions and data flow
- Text label content
- The scientific meaning of the diagram
Practice 2: Be specific about your target conference style
Different academic conferences have distinct illustration styles. Clearly specify this in your prompt:
- NeurIPS Style: Minimalist, soft color palettes, plenty of white space.
- CVPR Style: Often uses blue and purple tones, focuses on visual hierarchy.
- ACL Style: Compact layouts, emphasizes text-based workflows.
Practice 3: Polish complex diagrams in stages
For complex diagrams with more than 10 modules, it's better to polish them step-by-step:
- Start by refining the overall layout and color scheme.
- Then, optimize the details of key sub-modules individually.
- Finally, adjust for global consistency.
Practice 4: Use Critic feedback to iterate and improve
The PaperBanana Critic agent provides specific suggestions for improvement. When making API calls, you can feed the previous round's Critic feedback into the next round's prompt:
Previous Critic feedback: "Module spacing inconsistent between layer 2 and layer 3.
Suggest uniform 20px spacing."
Please fix this in the next iteration.
Practice 5: Prioritize scientific accuracy above all else
The most important rule during polishing is that scientific accuracy can never be compromised. Always emphasize these points in your prompt:
- "Preserve all original information elements"
- "Do NOT add or remove any nodes/connections"
- "Keep all labels and annotations unchanged"
Practice 6: Build your own diagram style library
Save your favorite results as reference templates to use in future calls:
Polish this diagram in the same style as my previous figure
(soft pastels, flat icons, L-R flow, 20px spacing, sans-serif 10pt)
Nano Banana Pro Diagram Polishing FAQ
Q1: Does polishing change the scientific meaning of the original diagram?
No, it doesn't. PaperBanana's diagram polishing strictly follows the principle of "Preserve structure, optimize presentation." The Critic agent gives the highest weight (30%) to "Faithfulness" to ensure all information nodes, connections, and text labels remain intact. If you submit an architecture diagram with 8 modules, you'll get 8 modules back—no more, no less. When calling via the APIYI (apiyi.com) platform, you can add "Preserve all scientific content" to your prompt to further reinforce these constraints.
Q2: What's the cost per polish? What about the total for 3 iterations?
A single Nano Banana Pro API call costs $0.05 (via the APIYI apiyi.com platform). Completing a 3-round iterative refinement costs a total of $0.15. If you generate two different color schemes at once (e.g., one "Soft Tech Pastels" and one "Deep Academic"), the total cost is about $0.25–$0.30. Compared to the official price of $0.234 per call, using the APIYI platform can save you about 80%, making it perfect for bulk optimization before paper submissions. You can also visit Image.apiyi.com for a quick online experience with single diagrams.
Q3: Which input formats are supported for the original diagrams?
PaperBanana primarily takes original diagram information via text descriptions. You'll need to describe the following in your prompt:
- Diagram structure (modules, hierarchy, connections)
- Text label content
- Data flow
- The diagram's purpose and target conference
For existing image files, you can use Nano Banana Pro's Image-to-Image mode by uploading the original diagram and attaching your polishing instructions.
Q4: How do I pick the best color scheme for my paper?
Here's our recommended workflow:
- Check if your target journal or conference has specific formatting guidelines (some require grayscale).
- If there are no special requirements, use "Soft Tech Pastels" by default (it has a 65% usage rate in top-tier conferences).
- For data-heavy diagrams, consider the "Deep Academic" dark scheme.
- Generate 2–3 color schemes via the APIYI (apiyi.com) platform to compare and pick the best one.
Q5: Can I use the polished diagrams directly in my paper submissions?
Absolutely. The quality of diagrams polished by PaperBanana meets the publication standards of top-tier academic conferences. In PaperBananaBench benchmarks across 292 test cases, PaperBanana-polished diagrams achieved an overall win rate of 72.7%. They significantly outperformed original diagrams in conciseness (+37.2%), readability (+12.9%), and aesthetics (+6.6%). We still recommend a final manual check before submission to ensure all scientific information is perfectly accurate.
Nano Banana Pro Chart Polishing: Summary and Recommendations
PaperBanana's chart polishing feature represents a major paradigm shift in scientific illustration—moving from relying on personal aesthetic experience to systematic aesthetic optimization driven by data from top-tier conference papers.
Key Advantages Recap:
- Data-Driven: Aesthetic standards are derived from actual statistics of 292 NeurIPS papers, rather than subjective judgment.
- 4-Dimensional Optimization: Comprehensive, systematic polishing across layout, color schemes, fonts, and icons.
- 3-Round Iteration: A Visualizer + Critic refinement loop that simulates a professional designer's workflow.
- Preserving Scientific Integrity: Fidelity evaluation carries the highest weight, ensuring academic accuracy is never compromised.
- Extremely Low Cost: Using the APIYI platform, it's only $0.05 per call, or $0.15 for a full 3-round iteration.
We recommend trying out Nano Banana Pro's chart polishing capabilities via the APIYI (apiyi.com) platform. The platform offers a unified API interface and an online experience tool at Image.apiyi.com, letting you finish your first polished chart in just 5 minutes.
References:
- PaperBanana Project Homepage: dwzhu-pku.github.io/PaperBanana
- PaperBanana Paper: arxiv.org/abs/2507.07300
- NeurIPS 2025 Proceedings: neurips.cc
Author: APIYI Team | To discuss more technical details about Nano Banana Pro chart polishing, feel free to contact us via the APIYI (apiyi.com) platform for technical support and free test credits.
