Automatic Method Illustration Generation for AI Scientific Papers via Drawing Middleware Creation, Evolution, and Orchestration

Li, Zhuoling; Zhang, Jiarui; Hu, Ping; Kuen, Jason; Gu, Jiuxiang; Rahmani, Hossein; Liu, Jun

ECCV 2026

Automatic Method Illustration Generation for AI Scientific Papers via Drawing Middleware Creation, Evolution, and Orchestration

Zhuoling Li^1* Jiarui Zhang^1* Ping Hu² Jason Kuen³ Jiuxiang Gu³ Hossein Rahmani¹ Jun Liu¹

¹ Lancaster University, ² University of Electronic Science and Technology of China, ³ Adobe Research

§Method figures should be precise and editable.

Method illustrations carry the logic of a scientific paper, but creating them still takes hours or days of careful composition.

FigAgent mirrors how human authors work: study strong examples, preserve useful visual components as reusable drawing middlewares, adapt those middlewares as visual requirements evolve, and try multiple drawing paths before committing to a final composition.

Reuse

Distill drawing experience

Recurring visual components become parameterized Python functions that generate DrawIO XML.

Evolve

Adapt the visual vocabulary

Selection, mutation, and crossover extend middlewares to cover newly emerging concepts.

Delay premature choices

Candidate drawings are explored and evaluated for their long-term effect on the complete figure.

Method

§Five agents, one shared canvas.

The framework separates repository construction from paper-specific illustration generation while keeping the final artifact editable.

Overview of the FigAgent framework — FigAgent creates and evolves a reusable middleware repository, then orchestrates it to draw an illustration concept by concept.

Constructor

Builds the experience dataset and maintains the drawing middleware repository.

Parser

Extracts a research theme, core concepts, and their relationships as a concept graph.

Drawer

Retrieves, parameterizes, and orchestrates middlewares on the DrawIO canvas.

Evaluator

Produces revision feedback and quality scores for candidate canvas states.

Refiner

Adds connections, tunes spacing, and harmonizes the final visual style.

§A middleware can grow with the research community.

Run the three Python middlewares and inspect how creation, mutation, and crossover change the same visual component.

View the paper figure

Evolution of a drawing middleware from an identified component to new graph variants

Creation

create_layer_graph

A parameterized graph component created from an identified visual pattern.

Preparing Python runtime…

Loading interactive preview

Node shape Node count 6 Layout Edge width 2 Labels Container Sketch mode

Python

Gallery

§Qualitative comparisons and generated illustrations.

Compare FigAgent with representative baselines.

Qualitative comparisons between FigAgent and representative baselines, with additional generated method illustrations

§Cite FigAgent.

The citation will be updated when the official proceedings metadata becomes available.

@inproceedings{li2026figagent,
  title     = {Automatic Method Illustration Generation for AI Scientific Papers via Drawing Middleware Creation, Evolution, and Orchestration},
  author    = {Li, Zhuoling and Zhang, Jiarui and Hu, Ping and Kuen, Jason and Gu, Jiuxiang and Rahmani, Hossein and Liu, Jun},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2603.29590}
}