paper-banana

Generate publication-quality academic illustrations using PaperBanana multi-agent pipeline

Use this agent to critique and refine figure descriptions by comparing the generated image against the source content. The critic provides feedback and revised descriptions for iterative improvement in the PaperBanana pipeline. <example> Context: A diagram has been generated and needs quality review user: "Critique the generated diagram against the methodology" assistant: "I'll use the critic agent to review the diagram and suggest improvements." <commentary> The critic compares the generated image against source content and provides structured feedback. </commentary> </example> <example> Context: A plot needs review for data accuracy user: "Review the generated plot for data fidelity" assistant: "I'll use the critic agent to verify the plot's accuracy and suggest corrections." <commentary> The critic ensures plots accurately represent all data points and visual intent. </commentary> </example>

Use this agent to generate a detailed textual description of an academic figure (diagram or plot) based on the methodology/data and reference examples. This is the planning step that creates the initial figure description for the PaperBanana pipeline. <example> Context: Reference examples have been retrieved and planner needs to create description user: "Generate a detailed description for the target diagram" assistant: "I'll use the planner agent to create a detailed figure description based on the references." <commentary> The planner creates the initial detailed description that subsequent agents will refine. </commentary> </example> <example> Context: Pipeline state has references and needs planning user: "Plan the figure description for the plot" assistant: "I'll use the planner agent to generate the plot description using few-shot examples." <commentary> The planner uses retrieved references as few-shot examples to generate better descriptions. </commentary> </example>

Use this agent to retrieve relevant reference examples from PaperBananaBench dataset for few-shot learning in the PaperBanana pipeline. This agent selects the top 10 most relevant reference diagrams or plots from a candidate pool. <example> Context: The PaperBanana pipeline needs reference examples for few-shot learning user: "Retrieve reference examples for diagram generation" assistant: "I'll use the retriever agent to select relevant reference examples from the dataset." <commentary> The retriever agent searches the PaperBananaBench dataset for similar examples to guide generation. </commentary> </example> <example> Context: Starting the full pipeline and need to find similar plots user: "Find reference plots for statistical visualization" assistant: "I'll use the retriever agent to find the most relevant reference plots." <commentary> Retrieval is the first step in the full pipeline, providing few-shot examples to the planner. </commentary> </example>

Use this agent to refine and enrich a figure description with aesthetic details based on NeurIPS 2025 style guidelines. The stylist polishes the planner's output without changing semantic content. <example> Context: Planner has generated an initial description that needs aesthetic refinement user: "Polish the diagram description with NeurIPS style guidelines" assistant: "I'll use the stylist agent to add aesthetic details to the description." <commentary> The stylist refines visual attributes like colors, fonts, and layout without altering content. </commentary> </example> <example> Context: Plot description needs style enhancement user: "Apply style guidelines to the plot description" assistant: "I'll use the stylist agent to enrich the plot description with publication-ready styling." <commentary> The stylist ensures the description meets NeurIPS 2025 aesthetic standards. </commentary> </example>

Use this agent to generate actual images (diagrams via Gemini image API or plots via matplotlib code) from detailed textual descriptions. This is the visualization step in the PaperBanana pipeline. <example> Context: A detailed description has been generated and needs to be rendered as an image user: "Generate the diagram image from the description" assistant: "I'll use the visualizer agent to render the diagram using the Gemini image API." <commentary> The visualizer converts text descriptions into actual images using appropriate generation methods. </commentary> </example> <example> Context: A plot description needs to be converted to an actual matplotlib plot user: "Generate the statistical plot from the description" assistant: "I'll use the visualizer agent to generate matplotlib code and execute it." <commentary> For plots, the visualizer generates Python code and executes it to create the image. </commentary> </example>

Generate publication-quality academic illustrations (diagrams and statistical plots) for research papers using the PaperBanana multi-agent pipeline. Supports: (1) Diagrams - scientific method/architecture illustrations via Gemini image API, (2) Plots - statistical visualizations via matplotlib code generation. Use when user wants to create academic figures, conference paper illustrations, method diagrams, architecture diagrams, pipeline diagrams, statistical plots, bar charts, line charts, heatmaps, or NeurIPS/ICML-style academic visualizations. Also supports polishing existing figures and reading .tex/.md files to auto-generate paper illustrations.