How CRAFTER Turns AI‑Generated Research Figures into Editable SVGs

Problem

Current AI figure generators produce static PNG images that require manual adjustment of arrows, labels, colors, and layout before they can be used in a paper.

CRAFTER – Multi‑Agent Harness

CRAFTER closes the loop between generation, checking, and revision. It first parses the paper context and user intent, then generates multiple visual drafts. An image‑generation model creates raster outputs for each draft. After each generation step the system evaluates content accuracy, layout consistency, text readability, and visual defects, recording structured feedback. Based on this feedback it decides to accept, modify, or revert to a better version. Structured operations allow adding layout constraints, resizing specific elements, or suppressing recurring visual flaws.

Evaluation of CRAFTER

On the PaperBanana‑Bench and CRAFTBENCH benchmarks CRAFTER achieves the highest composite scores, ranging from 5.04 to 8.90. Ablation experiments show that removing any core mechanism (generation, checking, or structured operations) reduces the overall score.

CRAFTEDITOR – Raster‑to‑SVG Conversion

The conversion pipeline proceeds in three stages:

Canvas cleaning and element extraction.

For each element, generate a textual description and its position, then decide whether to treat it as a vector component or retain it as a raster asset.

Assemble an SVG skeleton and re‑check layout, text overflow, arrow endpoints, element overlap, and missing components.

CRAFTEDITOR supports four input conditions: text‑to‑image, mask completion, key‑element composition, and sketch‑conditioned generation. The accompanying CRAFTBENCH dataset contains 279 samples covering academic figures, posters, and infographics, providing diverse test cases.

SVG Conversion Results

Evaluated on 80 CRAFTER outputs, CRAFTEDITOR attains an overall score of 8.04 , outperforming AutoFigure‑Edit (6.91) and Edit‑Banana (3.69) across dimensions such as position, color, text, icons, arrows, and style.

Use Cases

Generate method diagrams directly from paper context and captions.

Complete blank regions in poster designs.

Refine sketches during design discussions.

Construct full figures from partial icons or elements.

Open‑Source Release

The project is released under an open‑source license. Code is available at https://github.com/HaozheZhao/Crafter. It requires configuring the SAM3 grounding server and is presented as a prototype rather than a turnkey web product; final figure quality still depends on author verification.