Let Your Data Build and Improve Itself

Why Dataset Construction Needs Agent Strategy

Dataset construction is no longer a single fixed path. Some tasks need image generation, some need editing or video synthesis, some need 3D reconstruction and Blender insertion, and paper reproduction needs source-backed methodology rather than ad hoc prompting.

WebSearch-Guided Reproducible Research Pipelines

The research-prior workflow turns a paper, topic, or generation request into a reproducible dataset-construction handoff. The active agent performs WebSearch and paper reading, then dynamically composes serial and parallel t2i, edit, i2v, and 3D routes while DataEvolver keeps source notes, trace events, gallery sheets, and research_prior.json attached for replay.

WebSearch-guided reproducible research pipelines The agent retrieves papers, official artifacts, code, and datasets, then records evidence, GT references, manifests, lineage, and research_prior.json.

Pipeline & Gallery Snapshots

Visual checkpoints for the edit route, universal paper synthesis, VLM review, and one-page handoff.

‹

Edit Route

Multi-case edit gallery

Parallel edit prompts, references, and generated outputs

Universal Mode

Four-paper case synthesis

Shared construction requirements distilled from related work

Review Loop

VLM review contact sheet

Candidate renders packaged with review-ready context

Handoff

One-page research handoff

Compact summary for downstream generation review

›

Single-Image Route Case

Robot Safety Signal: T2I → Edit → I2V

A compact three-stage example from one accepted image. The route starts with a text-to-image base render, promotes the selected frame through targeted image editing, then uses the edited result as the anchor for short motion synthesis.

1Base Image

1Edit Pass

1I2V Clip

MixStrategy

T2I Base

Initial robot render

Clean lab robot scene with a blue transparent cube held at the center.

Edit Pass

Safety signal added

Targeted edit adds a green status light and a front safety mark while preserving pose and scene layout.

I2V Output

Motion-ready result

The edited image becomes the visual anchor for a short image-to-video generation pass.

Serial / Parallel Strategy Orchestration

Use parallel search where alternatives are cheap, then switch to serial promotion once a stable image anchor is selected.

Parallel Explore base candidates

Fan out prompts, seeds, or style variants to quickly compare possible robot scenes.

Select Promote one anchor

Choose the frame with the clearest pose, object placement, and scene consistency.

Serial Apply targeted edit

Lock geometry and camera first, then add the safety indicator and status lights.

Serial Generate motion

Use the edited image as the input anchor for I2V so identity and visual edits carry forward.

Parallel T2I search→ Anchor selection→ Serial edit→ Serial I2V→ Replayable handoff

DataEvolver-Rotate: View-Controlled Rotation Editing

A benchmark dataset for rotation-conditioned image editing. Each sample pairs a canonical front-view image with a target view specified in natural language.

import json
from pathlib import Path
from PIL import Image

root = Path("dataset_scene_v7_full50_rotation8_...")
rows = []
with (root / "pairs/train_pairs.jsonl").open("r") as f:
    for line in f:
        rows.append(json.loads(line))

row = rows[0]
source = Image.open(root / row["source_image"]).convert("RGB")
target = Image.open(root / row["target_image"]).convert("RGB")
instruction = row["instruction"]

24 Structured Atomic Actions

In the 3D reconstruction route, the agent selects from a discrete, structured action space to address VLM-identified rendering issues. Each action targets a bounded Blender parameter.

+ 18 more actions — see scene_action_space.json

Built With

The models, frameworks, and traceability records powering DataEvolver's multimodal generation, 3D reconstruction, and research-prior workflows.

Citation

If you use DataEvolver or DataEvolver-Rotate in your research, please cite our work.

@misc{zhang2026dataevolverletdatabuild,
  title        = {DataEvolver: Let Your Data Build and Improve
                  Itself via Goal-Driven Loop Agents},
  author       = {Qisong Zhang and Wenzhuo Wu and Zhuangzhuang Jia
                  and Yunhao Yang and Huayu Zhang and Xianghao Zang
                  and Zhixiang He and Zhongjiang He and Kongming Liang
                  and Zhanyu Ma},
  year         = {2026},
  eprint       = {2605.01789},
  archivePrefix= {arXiv},
  primaryClass = {cs.AI},
  url          = {https://arxiv.org/abs/2605.01789}
}