Crafter: A Multi-Agent Harness for Editable Scientific Figure Generation from Diverse Inputs

Researchers introduce Crafter, a multi-agent system for generating publication-quality scientific figures from diverse inputs that generalizes across figure types without architectural changes. The work addresses a critical gap in automation tools by enabling editable SVG outputs and introduces CraftBench, a comprehensive benchmark for evaluating figure generation across multiple types and input conditions.

Scientific figure generation represents a significant bottleneck in academic publishing, consuming substantial researcher time despite being largely formulaic work. Crafter addresses this inefficiency through a novel multi-agent harness architecture that moves beyond single-task generators toward a unified system handling diverse figure types and input modalities simultaneously. The key innovation lies not in architectural complexity but in orchestrating specialized agents through a structured composition framework, enabling the system to handle localized errors that plague raster-based generators.

Existing automated figure generation tools have remained narrow in scope, each targeting specific figure types under constrained input conditions. This fragmentation forces researchers to use multiple tools or resort to manual creation. Crafter's generalization capability across figure types without retraining represents a methodological advance applicable beyond scientific visualization. The complementary CraftEditor system converts raster outputs to editable SVGs, addressing a critical usability gap where generated figures often require refinement that wasn't previously possible without reverting to source tools.

The introduction of CraftBench with human quality annotations establishes evaluation standards currently missing from the field. This benchmarking infrastructure matters as much as the system itself, enabling future work to measure progress against standardized criteria. The experimental results demonstrating substantial improvements over standalone generators and baseline agentic approaches validate the multi-agent harness pattern as architecturally superior for compositional tasks.

For the research community, this work promises tangible acceleration of paper preparation cycles. The open-source release and available benchmark position Crafter as a reference implementation that could influence future figure generation systems. Future development likely focuses on expanding supported figure types and improving SVG conversion fidelity.

• →Crafter's multi-agent harness generalizes across diverse figure types and input conditions without architectural changes, addressing a major limitation of existing narrow-purpose generators.
• →CraftEditor converts raster outputs to editable SVGs, enabling researchers to refine generated figures locally—solving a critical usability problem in automated figure generation.
• →CraftBench introduces the first standardized benchmark for scientific figure generation spanning multiple figure types and input conditions with human quality annotations.
• →The multi-agent composition approach represents a methodological pattern potentially applicable to other structured generation tasks beyond scientific visualization.
• →Open-source availability and benchmarking infrastructure position Crafter as a reference implementation that could influence future research in automated scientific document generation.