One Figure, Every Canvas: Agentic Pipeline for Editable Flowchart Relayout (arXiv 2610.06852)

Forum topic · 小凯 · 2026-10-07

Summary

A new arXiv paper (2610.06852) formulates aspect-ratio-adaptive flowchart relayout as a distinct computer vision task: given a raster flowchart from an ML paper and a target aspect ratio, generate a structure-faithful, hallucination-free, editable layout. Existing approaches fail in different ways: image-to-image models stretch boxes and reject extreme aspect ratios, text-to-image agentic systems hallucinate content, and parse-then-render systems misroute edges. The proposed solution is a three-stage agentic pipeline—Parse, Style, and Layout—where each stage pairs a main agent with a critic. Critics combine deterministic constraint checking with VLM visual feedback to explicitly verify connectivity and prevent silent breakage of edges. Outputs are draw.io-editable mxGraph XML. On a curated benchmark of 100 flowcharts across five aspect ratios, evaluated by Gemini 3.1 Pro and validated against human judgment, the method achieves 68.6% Content Fidelity versus 11.2–41.4% for prior work. Authors: Shih-Chen Tseng, Chih-Hsuan Chen, Ryan Yang, Hsi-An Chen, Chun-Wei Tuan Mu, Yu-Lun Liu.

Paper Overview

Field: Computer Vision (CV) Authors: Shih-Chen Tseng, Chih-Hsuan Chen, Ryan Yang, Hsi-An Chen, Chun-Wei Tuan Mu, Yu-Lun Liu Published: 2026-10-05 arXiv: 2610.06852

Motivation

Pipeline figures in ML papers must be adapted to many canvases: two-column papers, 16:9 slides, portrait posters, 1:1 social media covers, and 9:16 mobile previews. Each format imposes different aspect-ratio constraints on the same computation graph, and any silently broken connector distorts the method itself.

Task Definition

The paper defines aspect-ratio-adaptive flowchart relayout: given a raster flowchart and a target aspect ratio, produce a structure-faithful, hallucination-free, *editable* layout.

Why Prior Approaches Fail

Image-to-image models: stretch boxes and refuse extreme aspect ratios.
Text-to-image agentic systems: hallucinate content.
Parse-then-render systems: misroute edges.

Proposed Method

An agentic pipeline with three stages:

1. Parse 2. Style 3. Layout

Each stage pairs a main agent with a critic. Critics combine deterministic constraint checks with VLM visual feedback, explicitly verifying connectivity to prevent silent breakage of edges. The output is draw.io-editable mxGraph XML.

Results

On a curated benchmark of 100 flowcharts at five aspect ratios, evaluated with Gemini 3.1 Pro and validated against human judgment:

| Method | Content Fidelity | |---|---| | This work | 68.6% | | Prior work | 11.2%–41.4% |

Original Abstract

> Pipeline figures in ML papers must be repurposed across many canvases. We formulate aspect-ratio-adaptive flowchart relayout as a distinct task and propose an agentic pipeline with Parse, Style, and Layout stages, each pairing a main agent with a critic. Outputs are draw.io-editable mxGraph XML. On a curated benchmark of 100 flowcharts at five aspect ratios, our method reaches 68.6% Content Fidelity versus 11.2-41.4% for prior work.

*Auto-collected on 2026-10-07.*