TikZ vs Matplotlib vs AI: The Best Tools for Academic Paper Figures in 2026

If you have ever submitted a paper to a top venue, you know the drill. You write your training curves in matplotlib, draw your architecture diagram in draw.io, fine-tune the layout in Figma, and then spend an afternoon stitching everything together so the fonts, line weights, and color palettes at least appear consistent. For a single paper, you might touch three or four different tools — each with its own export quirks, coordinate systems, and learning curve.

This tool fragmentation has been a pain point for researchers for over a decade. But the landscape shifted dramatically in 2025 and 2026. AI-powered figure generators moved from novelty demos to genuinely useful research tools, and the question is no longer "should I learn TikZ or matplotlib?" but rather "where does AI fit into my figure workflow?"

This article is an honest, tool-by-tool comparison. We cover code-based tools (TikZ, matplotlib, ggplot2), GUI-based tools (draw.io, Figma, Illustrator, Excalidraw), and the new wave of AI-assisted generators (BioRender, FigureLabs, PaperBanana). We evaluate each on learning curve, output quality, reproducibility, speed, and price — so you can pick the right combination for your next paper.

Key Takeaways

• No single tool covers every figure type in a research paper. The most productive researchers use 2–3 tools together, typically one for data plots and another for schematic diagrams.
• Code-based tools (TikZ, matplotlib) produce the highest-quality output but demand significant time investment and domain expertise. They remain the gold standard for reproducibility.
• GUI tools (draw.io, Figma) are fastest for one-off diagrams but lack data-driven capabilities and can produce inconsistent styling across figures.
• AI-powered tools have crossed the usefulness threshold in 2026. Multi-agent pipelines that generate, critique, and refine figures iteratively now produce publication-quality results for common diagram and plot types.
• The key differentiator among AI tools is quality assurance. Single-pass generators often require extensive manual correction, while iterative critique-and-refine approaches (like PaperBanana's 7-agent pipeline) converge on higher quality with less human intervention.

The Academic Figure Landscape in 2026

Academic figure creation tools fall into three paradigms, each with distinct trade-offs:

Code-based tools (TikZ, matplotlib, ggplot2) prioritize reproducibility and precision. You write a script, version-control it alongside your data, and regenerate identical figures on demand. The cost is a steep learning curve and slow iteration cycles.

GUI-based tools (draw.io, Figma, Illustrator, Excalidraw) prioritize speed and visual flexibility. You drag, drop, and see results immediately. The cost is that figures are not data-driven, hard to reproduce programmatically, and prone to inconsistent styling across a paper.

AI-assisted tools (BioRender, FigureLabs, PaperBanana) prioritize accessibility and workflow speed. You describe what you need — in text, data, or a rough sketch — and an AI system generates a publication-quality figure. The cost is reduced fine-grained control and dependence on the generator's capabilities.

Here is a quick overview:

Code-Based Tools

TikZ / PGFPlots

TikZ is the gold standard for LaTeX-native figures. If you have ever admired a paper where the diagram fonts perfectly match the body text, where every arrow is pixel-aligned, and where the color palette feels deliberate — chances are it was made in TikZ.

Pros:

• Native LaTeX integration. Fonts, math notation, and spacing match your document automatically. No post-processing required.
• Vector output with precise control. Every coordinate, every curve, every annotation is defined in code. You get exactly what you specify.
• Excellent for methodology diagrams. Flowcharts, neural network architectures, pipeline diagrams, and state machines are TikZ's sweet spot. The results look unmistakably professional.
• Large community and example library. Sites like texample.net and tikz.net host thousands of ready-to-adapt examples. The TeX Stack Exchange has detailed answers for almost any layout question.

Cons:

• The learning curve is brutal. The PGF/TikZ manual exceeds 1,300 pages. Even experienced LaTeX users report spending hours on layouts that would take minutes in a visual editor.
• No WYSIWYG feedback loop. The workflow is: write TikZ code, compile the document, check the PDF, adjust coordinates, recompile. This cycle is slow, especially for spatial layouts where you need to eyeball positions.
• Requires significant mental arithmetic. As Hacker News user semi-extrinsic noted in a widely-cited thread, TikZ demands that you "do significant mental arithmetic to place elements" — a cognitive overhead that compounds with diagram complexity.
• LLMs struggle with TikZ generation. The DeTikZify paper presented at NeurIPS 2024 demonstrated that even state-of-the-art language models have difficulty generating correct, compilable TikZ code from figure descriptions. This limits the potential for AI-assisted TikZ workflows.

Best for: Researchers with deep LaTeX expertise who need pixel-perfect diagrams that typographically match their document. If you are writing a theory-heavy paper with formal diagrams that will be referenced across multiple publications, the TikZ investment pays off.

Verdict: Produces the best-looking results for those willing to invest the time. But for a deadline-driven paper, the time cost can be prohibitive.

matplotlib + seaborn

matplotlib is the workhorse of scientific Python. If your paper has training curves, ablation studies, bar charts, or scatter plots, you are almost certainly using matplotlib — directly or through seaborn's higher-level API.

Pros:

• Deep Python ecosystem integration. pandas DataFrames, NumPy arrays, Jupyter notebooks — matplotlib fits naturally into the ML research workflow. You can go from experiment logs to publication figures in the same script.
• Unlimited customization via rcParams. Every visual aspect — font sizes, line widths, tick formatting, legend placement — can be configured programmatically. Nothing is locked behind a GUI.
• Community style libraries for academic use. The SciencePlots library provides journal-specific styles (IEEE, Nature, Science). The tueplots and mpl_sizes libraries provide exact sizing for NeurIPS, ICML, ICLR, and other conference formats.
• Reproducibility. Save your plotting script alongside your data and anyone can regenerate your figures. This is increasingly expected by reviewers and a requirement for some journals.

Cons:

• Default styles are not publication-ready. This is a known and long-standing issue. GitHub issue #19028 requesting better default styles for publication was closed as won't-fix. Out of the box, matplotlib figures look like homework, not like camera-ready papers.
• Verbose API for basic styling. Adjusting font sizes, removing spines, setting consistent margins, and aligning subfigure labels requires boilerplate code that is easy to get wrong. A "simple" bar chart can easily require 40–50 lines of styling code.
• LaTeX font matching is tricky. Enabling text.usetex=True for LaTeX-rendered text requires a working TeX installation, slows down rendering significantly, and introduces its own class of cryptic errors. Without it, fonts will not match your LaTeX document.
• Subfigure stitching causes inconsistencies. When you compose multiple subplots into a single figure (common for ablation studies), maintaining consistent font sizes, line weights, and label positions across subplots requires careful manual coordination.

Best for: Data-heavy papers with many experimental plots — training loss curves, performance comparisons, ablation grids, confusion matrices. If your paper lives in Jupyter notebooks, matplotlib is the natural choice.

Verdict: The undisputed workhorse for data plots. But expect to spend significant time on styling if you care about visual quality. Pair it with SciencePlots or tueplots to save hours.

ggplot2 (R)

ggplot2 deserves a brief mention, especially for researchers in statistics, computational biology, and social sciences. Its "grammar of graphics" philosophy produces elegant, well-structured visualizations with relatively concise code. The defaults are more publication-friendly than matplotlib's, and packages like cowplot and patchwork make multi-panel figures straightforward.

The primary limitation for ML/AI researchers is ecosystem lock-in: ggplot2 lives in R, and most ML workflows are in Python. If your data pipeline is already in R, ggplot2 is excellent. If you are training models in PyTorch, the friction of moving data between languages usually is not worth it.

GUI-Based Tools

draw.io (diagrams.net)

draw.io is the most popular free diagramming tool among researchers, and for good reason: it is browser-based, requires no installation, and handles the most common diagram types well.

Pros:

• Free and accessible. No account required. Works in any browser. Also available as a desktop app and a VS Code extension.
• Good for standard diagram types. Flowcharts, system architecture diagrams, UML-style process flows — draw.io handles these quickly with its built-in shape libraries.
• Exports SVG and PDF. Vector exports integrate cleanly into LaTeX documents.

Cons:

• Performance degrades with complexity. Diagrams with more than a few dozen elements start to feel sluggish. Rearranging a complex methodology figure with many connections can be frustrating.
• Not data-driven. You cannot feed in a CSV and get a chart. Every element is manually placed.
• Limited academic-specific templates. The shape libraries lean toward software engineering (UML, network diagrams) rather than scientific figures. You will not find neural network layer blocks, experimental pipeline templates, or statistical plot elements.

Best for: Quick methodology diagrams and system architecture overviews when you need something "good enough" in 30 minutes. Many top-venue papers use draw.io figures, particularly for high-level pipeline overviews.

Figma

Figma has quietly become a favorite among design-savvy researchers, especially for collaborative projects and poster design.

Pros:

• Collaborative editing. Multiple co-authors can work on the same figure simultaneously. This is genuinely useful for multi-institution papers where people iterate on diagrams asynchronously.
• Vector-native with excellent alignment tools. Auto-layout, smart guides, and component systems make it easy to maintain visual consistency across a paper's figures.
• Excellent for polishing. If you have a rough figure from matplotlib or draw.io, Figma is ideal for the final visual pass — adjusting colors, alignment, and typography.

Cons:

• Not designed for data visualization. There is no built-in charting capability. You cannot import data and generate a plot.
• No LaTeX integration. Math notation requires workarounds (render LaTeX externally, import as SVG).
• Subscription cost for teams. The free tier is functional for individual use, but collaboration features require a paid plan.

Best for: Final visual polish on existing figures, collaborative poster design, and cases where you need precise control over aesthetics without writing code.

Adobe Illustrator

Illustrator is the industry gold standard for vector graphics, and many journals' production teams use it for final figure preparation.

Pros:

• Total design control. There is nothing you cannot do in Illustrator. Every path, gradient, transparency, and text element is fully adjustable.
• Industry-standard format support. Native AI files, EPS, PDF — whatever a publisher requires, Illustrator can produce it.

Cons:

• Expensive. At $22/month (or more as part of Creative Cloud), it is a significant cost for a tool you might use only during paper submission.
• Steep learning curve. Illustrator is a professional design tool. Using it effectively for scientific figures requires non-trivial skill.
• Not data-driven. Like Figma, there is no way to generate plots from data. It is a polishing tool, not a creation tool for data visualizations.

Best for: Camera-ready figure preparation when a publisher or advisor demands pixel-perfect visual quality. Some researchers use it exclusively for the final pass before submission.

Excalidraw

Excalidraw has gained popularity for its intentionally informal, hand-drawn aesthetic.

Pros:

• Minimal and fast. Opens instantly in a browser. No configuration. Start drawing immediately.
• Good for thinking. The hand-drawn style encourages sketching ideas without worrying about polish. Useful in the early stages of figure design.

Cons:

• Drawing tools are limited. As noted by Hacker News users, Excalidraw's toolset is deliberately simple. Complex diagrams quickly become unwieldy.
• Too informal for publication. The hand-drawn aesthetic, while charming, is not appropriate for most camera-ready submissions. Reviewers may perceive it as low effort.

Best for: Whiteboard-style brainstorming, quickly sketching an idea to share with collaborators, and planning figure layouts before committing to a more polished tool. Not suitable for final publication figures.

AI-Powered Tools

The AI-assisted figure generation space matured significantly in 2025–2026. Three tools stand out for academic use, each targeting different niches.

BioRender

BioRender has become the de facto standard for biological and life science illustrations, with over 50,000 pre-designed scientific icons and a drag-and-drop interface.

• Strengths: Massive library of biology-specific icons (cells, organs, lab equipment, molecular structures). The Anthropic/Claude partnership has added AI-assisted layout suggestions. Institutional site licenses mean many university researchers already have access.
• Best for: Biology, biochemistry, medical research, and life sciences. If you need a cell signaling pathway or a CRISPR mechanism diagram, BioRender is the clear choice.
• Price: $33/month for individual researchers. Many universities provide institutional licenses.
• Limitation for CS/ML researchers: The icon library is biology-centric. You will not find transformer attention blocks, GPU cluster diagrams, or training pipeline templates. There is no data plotting capability. For computer science and machine learning papers, BioRender is not the right tool.

FigureLabs

FigureLabs offers a versatile set of AI capabilities: sketch-to-figure conversion, text-to-figure generation, figure refinement, and automatic vectorization.

• Strengths: Broad capability set. The sketch-to-figure mode is genuinely useful — photograph a whiteboard drawing, and FigureLabs produces a clean, publication-style version. The vectorization feature converts raster images to SVG.
• Free plan: 400 credits per month, which covers roughly 20–40 figure generations depending on complexity.
• Limitation: Generation is single-pass. The AI produces one result per request, and if the output is not quite right, your options are to try a different prompt or manually edit the result. There is no built-in quality critique or iterative refinement loop.

PaperBanana Paper Visualizer

PaperBanana takes a different architectural approach: instead of a single model generating a figure in one pass, it uses a 7-agent pipeline where specialized AI agents collaborate and critique each other's work.

• The pipeline: Retriever (searches NeurIPS/ICLR figure databases for style references) → Planner (structures the visual layout) → Stylist (optimizes colors, fonts, and typography) → Visualizer (generates the image or matplotlib code) → Critic (evaluates across four dimensions: Faithfulness, Conciseness, Readability, Aesthetics) → Vanilla (produces a clean baseline) → Polish (final refinements). The Critic agent can trigger additional iterations, so the system self-improves before you see the result.
• Dual mode: Diagrams from text descriptions (methodology figures, architecture diagrams, pipeline overviews) and data plots from raw data (the system generates and executes matplotlib code, so the output is both a rendered plot and reproducible source code).
• Multi-round AI critique: This is the key differentiator. The Critic agent evaluates each iteration against explicit quality dimensions and provides structured feedback that drives the next iteration. In practice, this means the third iteration is substantially better than the first — a dynamic that single-pass tools cannot replicate.
• Reference-driven generation: The Retriever agent searches figures from top-venue papers to inform style decisions. This helps outputs follow established academic visual conventions rather than defaulting to generic styles.
• Transparency for plots: In plot mode, the generated matplotlib code is visible, editable, and reproducible. You are not locked into an opaque image — you can take the code and customize it further.
• Pricing: Free first draft. Credits for multi-round refinement.
• Best for: CS/ML researchers who need methodology diagrams or statistical plots with built-in quality assurance. The iterative critique loop is particularly valuable when you are unsure exactly what you want — the system explores and converges rather than requiring you to specify everything upfront.

Try it: PaperBanana Paper Visualizer

Decision Matrix: Which Tool Should You Use?

Here is a practical decision tree based on what you need to create:

• Need a data plot from experimental results? Use matplotlib if you want full manual control or PaperBanana plot mode if you want AI-assisted generation with reproducible code output.
• Need a simple flowchart or architecture diagram? Use draw.io. It is free, fast, and good enough for most high-level overviews.
• Need a complex methodology diagram matching your paper's visual style? Use TikZ if you have LaTeX expertise and time, or PaperBanana diagram mode if you want AI assistance with iterative refinement.
• Need biological or medical illustrations? Use BioRender. Nothing else comes close for life science iconography.
• Need to polish an existing figure for camera-ready submission? Use Figma (free, collaborative) or Adobe Illustrator (maximum control).
• Need reproducible plots matching a specific conference format? Use matplotlib + SciencePlots or matplotlib + tueplots. These libraries provide exact sizing and styling for NeurIPS, ICML, ICLR, and major journals.
• Need to quickly sketch an idea for collaborators? Use Excalidraw. Do not overthink it.

The Hybrid Approach

The most productive researchers do not commit to a single tool. They use 2–3 tools together, each for what it does best. Here are two common workflows we see:

Workflow A (AI-first for diagrams, manual for plots):

1. Generate an initial methodology diagram with PaperBanana Paper Visualizer using text from your methods section
2. Fine-tune colors and alignment in Figma
3. Write matplotlib scripts with SciencePlots for all data figures
4. Compose everything in LaTeX with consistent sizing

Workflow B (Manual-first with AI assists):

1. Write all data plots in matplotlib with tueplots for conference-specific sizing
2. Use PaperBanana for architecture diagrams and pipeline overviews where visual creativity matters more than data precision
3. Final polish pass in Illustrator or Figma for camera-ready submission

The key insight is that different figure types have different bottlenecks. Data plots are bottlenecked by styling and formatting — code-based tools with style libraries solve this well. Schematic diagrams are bottlenecked by creative layout and visual design — this is where AI tools provide the most value.

FAQ

Is draw.io good enough for a top-venue paper?

Yes, for certain figure types. Simple flowcharts, system architecture overviews, and pipeline diagrams made in draw.io appear regularly in NeurIPS, ICML, and ICLR papers. The key is keeping the design clean: use consistent colors, align elements carefully, and export as SVG or PDF for crisp rendering. Where draw.io falls short is complex methodology figures with many interconnected elements — the tool becomes cumbersome, and the results can look generic.

Should I learn TikZ as a PhD student?

It depends on your field and how many papers you expect to write. If you are in a theory-heavy area where you will produce 5+ papers with formal diagrams (state machines, proof structures, algorithm visualizations), learning TikZ pays dividends over your PhD. The figures will match your LaTeX typography perfectly, and you can build reusable templates.

If you are in an applied ML area where most figures are training curves and architecture overviews, the TikZ time investment is harder to justify. You might spend 20 hours learning TikZ for a result you could achieve in 2 hours with draw.io or an AI tool. Consider your opportunity cost.

Can AI tools replace manual figure creation entirely?

Not yet, and probably not in the near term for all cases. AI tools in 2026 excel at two specific tasks: generating schematic diagrams from text descriptions (methodology figures, architecture overviews) and producing standard data plots from raw data. For these common figure types, AI can save hours of work.

Where AI still falls short: highly custom visualizations with unusual layouts, data-dense figures requiring precise annotation placement, domain-specific conventions that the training data may not cover well, and figures that need to match an existing paper's exact style. For these cases, manual tools — or a hybrid approach — remain necessary.

What do top ML researchers actually use?

Based on publicly shared workflows and paper source files: matplotlib (with custom style files) for data plots is nearly universal. For schematic diagrams, the landscape is more diverse — TikZ for the LaTeX-proficient, draw.io for quick drafts, Keynote/PowerPoint (surprisingly common for simple block diagrams), and increasingly AI tools for initial drafts. Figma and Illustrator appear in the polishing stage. Most researchers express frustration with their figure workflow regardless of the tools they use — the "figure tax" is a widely shared pain point.

Are AI-generated figures acceptable at NeurIPS/ICML?

Yes, for schematic diagrams and data plots. The major ML venues have not prohibited AI-generated figures. The NeurIPS paper checklist asks authors to disclose AI tool usage, which is straightforward to comply with. Publisher AI policies (from Springer Nature, Elsevier, IEEE, etc.) primarily target AI-generated photorealistic images and prohibit AI-generated content that could be mistaken for real photographs. Schematic diagrams, flowcharts, and data visualizations — the types of figures these tools produce — are not the focus of these restrictions. That said, always check the specific submission guidelines for your target venue, as policies continue to evolve.

Conclusion

There is no single "best tool for scientific figures" — the answer depends on what kind of figure you need, how much time you have, and what tools you already know. But the landscape in 2026 is meaningfully better than it was even two years ago.

For data plots, matplotlib with a style library (SciencePlots, tueplots) remains the most reliable path to reproducible, publication-quality results. For schematic diagrams, the calculus has shifted: AI-powered tools now produce results that rival hand-crafted figures, especially when the AI system includes iterative quality critique rather than single-pass generation.

The multi-agent approach — where AI agents generate, evaluate, and refine figures through multiple iterations — represents a genuine step forward. It mirrors how human designers actually work: draft, critique, revise. This is why a system like PaperBanana's 7-agent pipeline, with its explicit Critic agent evaluating Faithfulness, Conciseness, Readability, and Aesthetics, consistently outperforms single-shot generators. The first draft is a starting point, not the final product.

Whatever tools you choose, the goal remains the same: figures that communicate your research clearly, look professional, and do not consume weeks of your time. If your current workflow involves four different tools, inconsistent styling, and late-night font debugging — it might be worth trying a different approach.

Try PaperBanana Paper Visualizer for your next paper and see how the multi-agent pipeline handles your methodology diagrams and data plots.

---

Want to learn more about AI-assisted scientific illustration? Read our guides on practical tips from a PhD researcher and how PaperBanana's multi-agent architecture works.