diagram-gen

Academic Diagram Generation

Generate NeurIPS-quality methodology diagrams or statistical plots from paper text.

Available Tools

Workflow

Follow these steps in order:

Step 0 — Optimize Inputs (optional)

Call optimize_input with the raw methodology text and figure caption. This:

1. Structures the methodology into diagram-ready format (components, flows, groupings)
2. Sharpens a vague caption into a precise visual specification

Recommended for long or complex methodology text, or vague captions. Use the optimized outputs as inputs to plan_diagram.

Step 1 — Plan

Call plan_diagram with the methodology text, figure caption, and diagram type. This single tool call:

1. Retrieves the most relevant reference examples from the curated set using a specialized retriever prompt
2. Loads their images and passes them to the VLM for visual in-context learning
3. Generates a comprehensive textual description using a dedicated planner prompt
4. Refines the description with NeurIPS-quality aesthetic guidelines via a stylist prompt
5. Recommends an aspect ratio based on content structure

You receive back an optimized, publication-ready description and a recommended aspect ratio.

Do NOT attempt to write the diagram description yourself. The plan_diagram tool produces significantly better descriptions because it uses reference images and dedicated prompts.

Step 2 — Generate

Call generate_image with the description returned by plan_diagram.

• For methodology diagrams: diagram_type = "methodology" (default)
• For statistical plots: diagram_type = "statistical_plot" and include raw_data
• Pass the aspect_ratio recommended by plan_diagram (e.g., "16:9", "4:3")

Step 3 — Critique & Refine (max 3 rounds)

Call critique_image with the generated image, the description, source text, and caption.

• If needs_revision is true: use the revised_description from the critique, then go back to Step 2.
• If needs_revision is false: the image is publication-ready. Done.

Repeat at most 3 total iterations.

Aspect Ratios

Supported: 1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9

Guidelines:

• Wide (16:9, 21:9): Left-to-right pipelines, sequential flows, encoder-decoder architectures
• Tall (2:3, 9:16): Top-to-bottom hierarchies, deep stacks, vertical tree structures
• Square-ish (1:1, 4:3, 3:4): Balanced architectures, grid layouts, multi-panel diagrams

Example Interaction

User: "Generate a methodology diagram for this paper: [text]"

Agent steps:
  1. plan_diagram(source_context=..., caption=..., diagram_type="methodology")
     → receives optimized description + recommended aspect ratio
  2. generate_image(description=<optimized_description>, diagram_type="methodology", aspect_ratio="16:9")
     → receives image path
  3. critique_image(image_path=..., description=..., source_context=..., caption=...)
     → if needs_revision: update description → generate_image again
  4. Reply with the final image path

Important Notes

• Always call plan_diagram first — it handles retrieval, planning, and styling in one step with visual in-context learning from real reference diagrams.
• Never use hex codes, pixel dimensions, or CSS values in descriptions — they render as garbled text in generated images.
• Never fall back to matplotlib or LaTeX for methodology diagrams — always use the image generation model via generate_image.
• For statistical plots, generate_image will automatically generate and execute matplotlib code.
• Pass user_feedback to critique_image if the user has specific comments about the generated image.