Feedback-driven data construction for text-rich image generation

DataEvolver Self-Evolving Multi-Agent Data Construction for Text-Rich Image Generation

Policy evolution. Turn rejection patterns into reusable guidance for the next construction round.

Four cooperative agents. Retriever, Verifier, Critic, and Generator form a closed feedback loop.

No model training required. The construction policy evolves entirely in natural-language space.

Strong transfer. The resulting data improves multiple text-rich image generators under matched budgets.

arXiv PDF Code

Data evolving from scattered cubes through a central crystal into structured visual panels

Siyu Yan^1,2,*, Yizhen Gao^1,*, Yilin Wang¹, Dongxing Mao¹, Alex Jinpeng Wang^1,†

¹Central South University ²The Hong Kong University of Science and Technology

^* Equal contribution · ^† Corresponding author

Overview

From static filtering to a learning construction process

Rejected data should improve the next round—not disappear.

Text-rich image generation demands realistic images, legible text, semantic alignment, and coherent layouts at the same time. Conventional crawl–filter–freeze pipelines inspect candidates once, discard failures, and freeze the accepted dataset.

DataEvolver treats construction as an evolving policy. Verification outcomes become semantic feedback that reshapes retrieval and targeted generation, allowing each round to learn from the failure modes observed in the previous one.

Overview of DataEvolver and OCR-F1 scaling results

DataEvolver overview. Rejection patterns are converted into feedback for iterative construction-policy updates.

Approach

A self-evolving multi-agent loop

Four agents turn verification signals into better data.

At round t, DataEvolver combines retrieval policy, generation-prompt policy, and an experience library of high-value feedback. The Critic updates this policy in natural language, without gradient updates or additional training.

DataEvolver closed-loop multi-agent construction pipeline

Closed-loop construction. Retrieval, verification, criticism, and targeted generation continuously refine the data policy.

Collect

Retriever

Builds a topic-aware candidate pool from an adaptive, policy-conditioned query space.

Diagnose

Verifier

Scores image quality, OCR quality, and semantic consistency while recording rejection causes.

Evolve

Critic

Turns round-level statistics and failures into natural-language construction guidance.

Complete

Generator

Targets under-covered topic regions and admits only synthetic samples that pass verification.

Evolution

Policy evolution in action

The system remembers why data failed.

When a construction round becomes duplicate-heavy, the Critic summarizes the pattern and requests more diverse layouts, colour schemes, writing styles, issuers, and sources. The resulting feedback becomes part of the next round’s policy memory.

278 → 148 deduplication failures after the policy update

Example of a Critic-guided DataEvolver policy update

Results

Stronger data under matched budgets

Consistent OCR gains that transfer across generators.

+85.3% OCR-F1 on TextScenesHQ at 0.75M scale

+35.3% OCR-F1 on LongTextBench at 0.75M scale

2× generator families showing transferable gains

Main results. DataEvolver improves OCR-oriented metrics while retaining competitive semantic alignment and visual quality.

Examples

Qualitative comparison

More legible text, stronger alignment, cleaner layouts.

Models trained on DataEvolver data produce text that is easier to read, more faithful to the prompt, and more consistent with the intended visual layout than fixed-dataset baselines.

Qualitative comparison of text-rich image generation models

BibTeX

@article{yan2026dataevolver,
  title   = {DataEvolver: Self-Evolving Multi-Agent Data Construction
             for Text-Rich Image Generation},
  author  = {Yan, Siyu and Gao, Yizhen and Wang, Yilin and
             Mao, Dongxing and Wang, Alex Jinpeng},
  journal = {arXiv preprint},
  year    = {2026}
}