How Nano Banana Pro Redefines AI‑Generated Scientific Visualizations

Introduction

Creating clear, professional figures that meet top‑conference aesthetics is essential for conveying complex model architectures and demonstrating scholarly rigor. Traditional drawing tools often become a “time black hole” because of high time costs and steep aesthetic thresholds.

Dual Application of Generative AI in Research Drawing

Using Nano Banana Pro as a case study, two core uses are explored: (1) rapid prototyping to break creative deadlocks, and (2) converting literature text into intuitive logical diagrams, aiming for a qualitative leap in drawing efficiency through end‑to‑end visualisation.

Precise Replication of Model Architecture

The methods section of a paper often requires turning linear textual descriptions into a clear two‑dimensional topology. The DeepSeek‑V3.2 DSA architecture is used to demonstrate how logical deconstruction enables visual reproduction of a complex sparse‑attention mechanism.

MAIN ARCHITECTURE (DeepSeek‑V3.2 DSA)
1. Inputs:
   Query token $h_t \in \mathbb{R}^d$
   Preceding context tokens $h_s \in \mathbb{R}^d$
2. Lightning Indexer Module:
   Projections: Maps $h_t \to q_{t,j}^I, w_{t,j}^I$ and $h_s \to k_{s}^I$.
   Scoring: Computes index scores $I_{t,s}=\sum w \cdot \text{ReLU}(q^I \cdot k^I)$, optimized for low‑cost FP8 computation.
3. Fine‑Grained Selector:
   Filtering: Identifies indices of Top‑k scores from $I_{t,:}$.
   Retrieval: Fetches specific Key‑Value entries $\{c_s\}$ for selected indices.
4. Core Backbone (Attention):
   MLA Architecture: Instantiated based on Multi‑Head Latent Attention.
   MQA Mode: Latent KV vectors shared across all query heads.
   Sparse Computing: Attends $h_t$ only to retrieved $\{c_s\}$, reducing complexity from $O(L^2)$ to $O(Lk)$.
   DATA FLOW: $h_t, h_s \rightarrow$ Lightning Indexer $\rightarrow$ Scores $I$ $\rightarrow$ Top‑k Selector $\rightarrow$ Sparse KV indices $\rightarrow$ MLA (MQA Mode) $\rightarrow$ Output $u_t$

Feeding the structured text into Nano Banana Pro produces a diagram that faithfully reproduces the “index‑filter‑compute” cascade, visually demonstrates how the Lightning Indexer reduces computational complexity, and automatically aligns visual style with academic norms (low‑saturation pastel palette, flat design, uniform line width, aligned modules).

Multidimensional Presentation of Abstract Concepts and Data

Concept visualization: Inputting the closed‑loop logic “perception‑decision‑execution” for embodied intelligence yields a flat, linear hierarchy diagram that captures temporal logic and translates the obscure “multimodal interaction” text into a colorful, modular flow, demonstrating that high‑quality concept figures can be generated without design expertise.

Data chart recreation: Providing a raw table from a paper and a brief style prompt enables Nano Banana Pro to quickly generate a pastel‑colored, proportionally accurate clustered bar chart. The tool automatically groups multidimensional data, removes redundant grid lines, and converts raw numbers into an intuitive trend visualization without loss of information.

Prompt Template Reference

You are an expert ML illustrator.
Draw a clean, NeurIPS/ICLR‑style scientific figure using Nano Banana Pro.
GOAL: Create a professional, publication‑quality diagram that exactly follows the structure and logic provided in the MODULE LIST below.
GLOBAL RULES:
- Flat, clean NeurIPS style (no gradients, no gloss, no shadows)
- Consistent thin line weights
- Professional pastel palette
- Rounded rectangles for blocks
- Arrows must clearly indicate data flow
- No long sentences, only short labels
- Keep spacing clean and balanced
LAYOUT:
- Horizontal left → right layout (recommended)
- Or vertical top → bottom if modules are inherently sequential
MODULE LIST (FILL THIS WITH YOUR PAPER'S CONTENT):
1. Input(s):
   - [Your input items]
2. Preprocessing / Encoding / Embedding:
   - [Your modules]
3. Core Architecture / Stages / Blocks:
   - [Your modules in exact order]
4. Special Mechanisms (optional):
   - [Attention / memory / routing / dynamic paths]
5. Output Head:
   - [Your output block]
NOTES (Optional but useful):
- Specify any required two‑branch or multi‑branch flow
- Specify “A and B must merge here”
- Specify “keep this as a single tall block with submodules”
- If experimental plot → replace section above with structured numbers
STYLE REQUIREMENTS:
- NeurIPS/ICLR 2025 visual tone
- Very light background
- Text left‑aligned inside blocks
- Arrows short and clean
- Use consistent vertical spacing
Generate the final diagram.