Hugging Face's logo

Datasets:
ruihangxu
/
RealManip-40K

Tasks:
Text-to-Image
Image-to-Image
Languages:
English
Size:
10K<n<100K
ArXiv:
Tags:
image-editing
3d-aware-image-editing
License:
Dataset card Data Studio Files
xet
 Community

Access RealManip-40K

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

RealManip-40K combines records governed by multiple upstream licenses. Some records are restricted to non-commercial or research use. Review LICENSE_DATA.md and the per-record license_ids before requesting access.

Log in or Sign Up to review the conditions and access this dataset content.

RealManip-40K

English | 简体中文

RealManip-40K is the real-world paired-image dataset introduced in PhyEdit: Towards Real-World Object Manipulation via Physically-Grounded Image Editing for training 3D-aware object manipulation models. Each record contains source and target RGB frames, object masks, estimated depth and camera parameters, representative object coordinates, bounding boxes, and editing prompts.

The nenhang/PhyEdit repository provides the accompanying helper tools and implementations of the ManipEval benchmark metrics.

This dataset repository contains 41,154 RealManip-40K training pairs and the 200-item ManipEval test split used to evaluate PhyEdit.

Dataset Splits

Split Dataset Records
train RealManip-40K 41,154
test ManipEval 200

Train Source Composition

Source dataset Collection Records
OpenUni Koala36M 20,660
OpenUni OpenS2V 2,155
UltraVideo Multiple collections 8,337
OpenVid-1M OpenVid and OpenVidHD 7,104
VIDGEN-1M Multiple collections 2,302
TrackingNet Training collections 506
GOT-10k Train and validation collections 90
Total 41,154

The train assets are content-addressed and deduplicated across records:

Modality Unique files Payload size (GiB)
RGB images 51,654 5.173
Depth maps 51,654 27.432
Object masks 95,247 0.198
Camera matrices 103,308 0.016
Total 301,863 32.819

The test split contains 2,235 unique assets with a 0.256 GiB payload:

Modality Unique files Payload size (GiB)
RGB images 400 0.0442
Depth maps 400 0.2103
Object masks 635 0.0015
Camera matrices 800 0.0001

Repository Layout 

metadata/train.jsonl             Public training records
metadata/schema.json             JSON Schema for each record
metadata/assets.jsonl            Deduplicated asset inventory and provenance
metadata/asset_checksums.jsonl   Per-asset SHA-256 and containing shard
metadata/shards.json             Tar sizes and SHA-256 checksums
metadata/test.json               ManipEval test records
metadata/test_schema.json        JSON Schema for test records
metadata/test_assets.jsonl       Test asset inventory and provenance
metadata/test_asset_checksums.jsonl
metadata/test_shards.json        Test tar sizes and SHA-256 checksums
metadata/license_registry.json   License identifiers and authoritative URLs
metadata/stats.json              Generated release statistics
data/images-*.tar                RGB image shards
data/depths-*.tar                Depth-map shards
data/masks-*.tar                 Mask shards
data/cameras-*.tar               Intrinsic/extrinsic matrix shards
data/test/*.tar                  ManipEval modality shards

Asset paths in metadata/train.jsonl are relative to the repository root, for example assets/images/54/...jpg. Camera intrinsics and extrinsics are stored as .npy files referenced by relative paths in the metadata.

Download and Extract

Install the Hugging Face Hub CLI, accept the access terms on the dataset page, authenticate locally, and download the snapshot:

hf download ruihangxu/RealManip-40K \
  --repo-type dataset \
  --local-dir RealManip-40K

cd RealManip-40K
python scripts/extract_shards.py --split train
python scripts/extract_shards.py --split test

The extractor rejects links and unsafe paths, verifies every tar and asset checksum, and safely resumes by verifying files that already exist. To extract only selected modalities, pass for example --modalities images masks.

The PhyEdit data loader resolves these paths directly when pointed at the public metadata:

export AGGREGATED_DATASET_ROOT=/path/to/RealManip-40K
export AGGREGATED_METADATA_PATH=/path/to/RealManip-40K/metadata/train.jsonl

Use /path/to/RealManip-40K/metadata/test.json as the benchmark metadata path for ManipEval sampling and evaluation.

The public test records use continuous identifiers: bench_index ranges from 0 to 199, and sample_id ranges from manipeval_0000 to manipeval_0199.

Data Format

Object-aligned fields are lists, and one record may manipulate multiple objects. resolution is [width, height] for the RGB frames. Bounding boxes use pixel coordinates. See metadata/schema.json for the complete field contract.

RGB images and masks retain their frame resolution. Depth maps are float32 NumPy arrays generated at an aspect-preserving resolution with long side 504; most 16:9 records are 280 x 504. Intrinsics are defined on the depth grid. When depth is resized to the RGB or training resolution, scale the focal lengths and principal point by the same horizontal and vertical factors. The alignment and 3D unprojection utilities used by PhyEdit are available in the PhyEdit repository. Intrinsics have shape 3 x 3, and extrinsics have shape 3 x 4.

Each record includes source_dataset, source_collection, source_record_id, and source_frame_ids for provenance. The license_ids list resolves through metadata/license_registry.json and may contain more than one applicable set of terms. ManipEval records also provide public benchmark identifiers and object depth-shift summaries used by the evaluation pipeline; see metadata/test_schema.json for their definitions.

Licensing and Responsible Use

There is no unified license for all records. See LICENSE_DATA.md before downloading or using the assets. Several source branches are restricted to non-commercial or research use, and all terms attached through license_ids apply simultaneously. The repository-level license: other tag does not replace those upstream terms.

The source videos are real-world data and may include identifiable people, logos, text, or other incidental content. The release has not been exhaustively audited for every possible privacy, demographic, or safety concern. It must not be used for identifying people or other harmful surveillance applications. For removal or metadata-correction requests, contact ruihangxu@zju.edu.cn with the relevant sample_id or source record identifier.

Depth and camera values are model estimates rather than sensor ground truth. The data-mining and VLM-filtering pipeline can also retain occasional imperfect masks, correspondences, or prompts. Users should account for this noise when training or evaluating systems.

The original helper scripts under scripts/ are available under the MIT License. That code license does not apply to dataset records, metadata content, or third-party assets.

Citation 

@misc{xu2026phyeditrealworldobjectmanipulation,
      title={PhyEdit: Towards Real-World Object Manipulation via Physically-Grounded Image Editing},
      author={Ruihang Xu and Dewei Zhou and Xiaolong Shen and Fan Ma and Yi Yang},
      year={2026},
      url={https://arxiv.org/abs/2604.07230},
}
Downloads last month
-
Total file size:
36 GB

Models trained or fine-tuned on ruihangxu/RealManip-40K

Paper for ruihangxu/RealManip-40K