brycewang-stanford/latex-drawing-collection

LaTeX Drawing Collection

A skill providing ready-to-use LaTeX drawing examples and guidance for creating publication-quality scientific figures using TikZ, PGFPlots, and related packages. Based on awesome-latex-drawing (2K stars), this skill covers Bayesian networks, tensor decompositions, neural architectures, time series visualizations, and more.

Overview

High-quality figures are essential for effective scientific communication. While external tools like Matplotlib or Inkscape can produce figures, native LaTeX drawings offer superior integration with the document, consistent typography, vector-quality output at any resolution, and automatic style matching with the surrounding text.

This skill equips the agent with knowledge of 30+ LaTeX drawing patterns commonly used in academic publications. Each pattern includes the required packages, a description of the drawing approach, and guidance on customization for specific research contexts.

Essential Packages

The following LaTeX packages form the foundation for scientific drawing:

TikZ (tikz)

• The core drawing package for LaTeX, providing a programming interface for vector graphics
• Supports coordinate systems, transformations, path operations, and decorations
• Required for virtually all complex scientific diagrams
• Load with: \usepackage{tikz} and relevant libraries via \usetikzlibrary{...}

PGFPlots (pgfplots)

• Built on TikZ for creating publication-quality data plots
• Supports 2D and 3D plots, error bars, fill areas, and custom markers
• Handles axis formatting, legends, and annotations
• Load with: \usepackage{pgfplots} and \pgfplotsset{compat=1.18}

TikZ Libraries

• arrows.meta - customizable arrowhead styles
• positioning - relative node placement (above=of, right=of)
• fit - bounding boxes around groups of nodes
• matrix - grid-based node layouts
• decorations.pathreplacing - braces, zigzag, snake decorations
• calc - coordinate arithmetic
• backgrounds - layered drawing with background regions

Bayesian Network Diagrams

Bayesian networks are among the most common diagrams in probabilistic modeling papers:

Node Styles

• Observed variables: filled circles or shaded nodes
• Latent variables: open (unfilled) circles
• Hyperparameters: small solid dots or fixed-value nodes
• Plates: rounded rectangles indicating repetition with index labels

Construction Approach

• Define node styles at the beginning of the tikzpicture environment
• Place nodes using relative positioning for maintainable layouts
• Draw directed edges with arrow styles indicating conditional dependencies
• Add plate notation around repeated variable groups
• Label edges with conditional probability annotations when needed

Common Patterns

• Latent Dirichlet Allocation (LDA) plate diagram
• Hidden Markov Model (HMM) chain structure
• Variational autoencoder (VAE) graphical model
• Gaussian mixture model (GMM) with plate notation
• Deep generative model hierarchies

Tensor and Matrix Diagrams

For linear algebra and tensor decomposition papers:

Tensor Representations

• Matrices as 2D grids with element shading
• Third-order tensors as 3D cubes with visible faces
• Tensor networks as connected node diagrams
• Factor matrices as thin rectangular blocks

Decomposition Visualizations

• CP decomposition: tensor equals sum of rank-one components
• Tucker decomposition: core tensor multiplied by factor matrices
• Tensor train: chain of connected 3D cores
• Matrix factorization: large matrix as product of thin matrices

Neural Network Architectures

For deep learning and machine learning papers:

Layer Representations

• Fully connected layers as columns of nodes with all-to-all connections
• Convolutional layers as stacked feature map grids
• Attention layers as matrix operation diagrams
• Recurrent connections as self-loops or unrolled sequences

Architecture Patterns

• Encoder-decoder structures with bottleneck
• Skip connections and residual blocks
• Multi-head attention mechanisms
• Transformer block diagrams

Time Series and Spatiotemporal Plots

For data analysis and forecasting papers:

Time Series Elements

• Line plots with confidence bands using PGFPlots fill between
• Missing data indicators with dashed segments
• Multi-variate time series as stacked or aligned panels
• Seasonal decomposition as vertically arranged subplots

Spatiotemporal Grids

• Heatmaps using TikZ matrix with color-coded cells
• Geographic grids with observation points
• Temporal slices showing spatial evolution

Customization Guidelines

When adapting templates for specific publications:

• Match the font size to the document class (typically 8-10pt for figure labels)
• Use consistent color schemes that work in both color and grayscale
• Align arrow styles across all figures in the paper
• Keep node sizes proportional to their importance in the diagram
• Add descriptive labels rather than relying solely on mathematical notation
• Test figures at the target column width before finalizing

Integration with Research-Claw

This skill supports the Research-Claw writing workflow:

• Generate LaTeX drawing code from verbal descriptions of desired figures
• Adapt existing templates to match specific research contexts
• Debug TikZ compilation errors and suggest fixes
• Recommend appropriate diagram types for different data structures
• Produce standalone compilable .tex files for figure testing

Best Practices

• Always use relative positioning instead of absolute coordinates for maintainability
• Define reusable styles at the document or figure level to ensure consistency
• Compile figures as standalone documents first, then include in the main paper
• Use \footnotesize or \scriptsize for labels inside dense diagrams
• Export to PDF for vector quality and include via \includegraphics
• Keep TikZ code well-commented for future modifications by collaborators