Dataset Card for Omni3DEdit

Dataset Summary

Omni3DEdit is a unified benchmark for instruction-guided 3D editing with explicit edited-region annotations. It addresses the gap in prior benchmarks that rarely include explicit edited-region supervision or a standardized protocol for assessing whether models edit the intended parts and preserve the rest.

The dataset contains 128,906 paired source-target 3D assets spanning five subsets that cover pose changes, rigid part addition and removal, articulated joint-state changes, region-targeted generative modifications, and material edits. Each sample provides a source asset, a target asset, a natural-language instruction, an edited-region annotation, lightweight edit metadata, and provenance information.

Supported Tasks and Leaderboards

This dataset is designed for evaluating instruction-guided 3D editing models. Key evaluation dimensions include:

• Edit Intent Alignment (Faithfulness): Whether the generated edit follows user intent described in natural language, measured by metrics like CLIP-T and DINO-I.
• Non-edited Region Preservation: Structural and visual stability in untouched regions, measured by Chamfer Distance (CD), mPSNR, mSSIM, and mLPIPS.
• Localized Edit Behavior (Locality): Whether the modified support aligns with target edit regions, evaluated via Region-F1.
• Cross-Scenario Generality: Unified protocol across multiple heterogeneous edit families.

Leaderboard: Official test split results are available on the project website.

Languages

The instructions are provided in English (en).

Dataset Structure

The repository is organized by edit families, separating the official evaluation test set from the training splits. Because of the large number of 3D assets, the .glb files and rendered images are compressed into .tar archives.

Directory Hierarchy

Omni3DEdit/
├── Articulation/
├── Material/
├── Part-Edit/
├── Pose/
├── Structure/         # (Example of a training subset folder)
│   ├── cond_img.tar   # Rendered reference images used for image-conditioned evaluation
│   ├── mask.tar       # Compressed explicit edited-region annotations (masks) in .glb format
│   ├── pair.csv       # Core metadata, instructions, edit operations, and asset IDs
│   └── raw.tar        # Compressed source and target 3D assets in .glb format
├── test/
│   ├── test.csv       # Metadata and instructions for the 500 official test pairs
│   └── test.tar       # Contains all source, target, and mask .glb files for evaluation
├── README.md
└── .gitattributes

File Contents

1. Training Subsets (e.g., Structure/, Pose/)

Inside each of the five edit family folders, you will find the following structure:

• pair.csv: The core metadata file containing mapping information, natural language instructions, edit operations, and corresponding asset IDs.

• raw.tar: The compressed archive containing the source and target 3D assets in .glb format.

• mask.tar: The compressed archive containing the explicit edited-region annotations (masks) in .glb format. (Note: The Pose subset does not contain localized edited-region annotations because pose transformations are inherently global)

• cond_img.tar: Rendered reference images used for image-conditioned evaluation.

2. Official Test Split (test/)

To facilitate standardized benchmarking, the 500 manually curated, human-verified test pairs are grouped together:

• test.csv: Contains the metadata, instructions, and file mappings specifically for the benchmark protocol.
• test.tar: A single archive containing all necessary .glb files (source, target, and masks) to run the official evaluation.

How to Load the Data

To use this dataset locally, we recommend cloning the repository via git-lfs and extracting the archives manually, or using the Hugging Face huggingface_hub Python library to download specific subsets.

Once extracted, you can load the metadata and map it to the local 3D assets using pandas:

import pandas as pd

# Load metadata for the Structure subset
df_structure = pd.read_csv("./Structure/pair.csv")

# Load metadata for the official test set
df_test = pd.read_csv("./test/test.csv")
print(df_test.head())

Dataset Creation

Curation Rationale

Most existing benchmarks only judge overall realism or global similarity, making it difficult to systematically assess where an edit occurs. Omni3DEdit standardizes heterogeneous data sources into a common GLB-based representation and provides exact edited-region annotations. This directly enables the evaluation of whether an instruction is faithfully realized while preserving unedited content.

Source Data

The dataset is constructed from diverse 3D assets across five edit families, totaling 128,906 pairs:

Edit Family	Pairs	Description	Source Assets
Pose	23,648	Pose changes from animated or rigged assets with frame pairs selected by deformation magnitude.	Objaverse
Structure	40,000	Rigid part addition and removal operations built from hierarchical part annotations.	PartNet
Articulation	6,875	Articulated joint-state changes with motion-aware edited regions derived from kinematic sweeps.	PartNet-Mobility
Part-Edit	26,407	Region-targeted generative modifications with explicit part targeting and realistic instructions.	PartNet, VoxHammer
Material	31,976	Appearance-only edits on texture or PBR materials while preserving underlying geometry.	3DCoMPaT++

Annotations

Each pair is standardized into a unified schema that includes an edited-region annotation (mesh-space faces/vertices plus projected per-view masks). For programmatic subsets (Structure, Articulation, Material), this annotation is exact by construction, providing a noise-free reference baseline for evaluation.

Considerations for Using the Data

Social Impact of Dataset

This benchmark aims to advance research in instruction-guided 3D editing, enabling more precise and controllable 3D content creation. Potential applications include virtual reality, gaming, 3D design, and robotics.

Discussion of Biases

The benchmark is constructed from available 3D assets and may inherit category imbalances and language artifacts from its upstream sources (e.g., Objaverse, PartNet). We mitigate this for evaluation through manual curation of the official test set.

Other Known Limitations

• Global Edits: The Pose subset lacks localized edited-region annotations because pose changes are inherently global transformations.

• Instruction Distribution: Because source-target pairs are largely constructed through programmatic transformations or foundation-model-based generation, the instruction distribution may not fully reflect real user requests.

• Scope: The benchmark currently focuses on object-centric edits and does not cover scene-level compositional editing.

Additional Information

Dataset Curators

The dataset was curated by Hongxing Fan, Haotian Lu, Rui Chen, Weibin Yun, Zehuan Huang, and Lu Sheng at Beihang University.

Licensing Information

Please refer to the project website for licensing details.

Citation Information

If you use this dataset, please cite the following paper:

@inproceedings{fan2026omni3dedit,
  title={OMNI3DEDIT: A Unified 3D Editing Benchmark with Region Annotations},
  author={Fan, Hongxing and Lu, Haotian and Chen, Rui and Yun, Weibin and Huang, Zehuan and Sheng, Lu},
  year={2026}
}

Contributions

Thanks to all contributors who made this dataset and benchmark possible.