| --- |
| language: |
| - en |
| license: other |
| license_name: worldcrafter-academic-research-only |
| license_link: LICENSE |
| pretty_name: WorldCrafter dataset |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: metadata/samples.jsonl |
| size_categories: |
| - 10K<n<100K |
| task_categories: |
| - image-to-video |
| - text-to-video |
| tags: |
| - worldcrafter |
| - camera-control |
| - video |
| - academic-research-only |
| - camera-pose |
| extra_gated_heading: Agree to access the dataset |
| extra_gated_description: >- |
| This dataset is restricted to non-commercial academic research at universities |
| and non-profit academic research institutions. Read the Dataset Access Agreement linked on this page, |
| provide your institution and research purpose, and accept the agreement. |
| Download access is granted automatically to your Hugging Face account upon submission. |
| extra_gated_button_content: Agree and access dataset |
| extra_gated_prompt: "Read the Dataset Access Agreement: https://huggingface.co/datasets/Drexubery/WorldCrafter/blob/main/LICENSE. By submitting, you accept its academic-research-only terms. Please consider starring https://github.com/TencentARC/WorldCrafter to support our open-source work; this is optional." |
| extra_gated_fields: |
| Full legal name: text |
| University or non-profit academic research institution: text |
| Department and academic position: text |
| Institutional email address: text |
| Institutional profile or research group URL: text |
| Research project and intended use: text |
| I have read and accept the WorldCrafter Academic Research Only Dataset License: checkbox |
| I will use the dataset solely for non-commercial academic research and not for commercial research or products: checkbox |
| I will not redistribute the dataset or share access and understand that each researcher must apply individually: checkbox |
| --- |
| |
| # WorldCrafter dataset |
|
|
| **High-quality videos with captions and camera annotations for camera-controlled video generation and WorldCrafter fine-tuning.** |
|
|
| [Project code](https://github.com/TencentARC/WorldCrafter) · [Project page](https://drexubery.github.io/WorldCrafter/) |
|
|
| If you find this release useful, please **[star WorldCrafter on GitHub](https://github.com/TencentARC/WorldCrafter)** to support our open-source work. A star is appreciated but is not required for access. |
|
|
| ## Preview |
|
|
| Most videos in the WorldCrafter dataset feature subject-following camera motion, helping models learn camera control for tracking and following subjects. |
|
|
| <video src="https://huggingface.co/datasets/Drexubery/WorldCrafter-preview/resolve/8738c69d65b3eac30aa500624db7006ba2f486d3/preview.mp4" poster="https://huggingface.co/datasets/Drexubery/WorldCrafter-preview/resolve/8738c69d65b3eac30aa500624db7006ba2f486d3/preview.jpg" controls autoplay muted loop playsinline width="100%"></video> |
|
|
| ## Access agreement |
|
|
| This dataset is available for **non-commercial academic research only**. |
| Read the [Dataset Access Agreement](https://huggingface.co/datasets/Drexubery/WorldCrafter/blob/main/LICENSE), |
| then sign in to Hugging Face, provide your institution and research purpose, |
| and accept the agreement using the access form on this page. Download access |
| is granted automatically to your Hugging Face account upon submission. |
|
|
| ## Contents |
|
|
| | Item | Count | |
| |---|---:| |
| | Unique source MP4 videos | 19,972 | |
| | Camera annotation JSON records | 20,434 | |
| | Valid target chunk entries | 77,642 | |
|
|
| Videos are deduplicated by their source-file path. Multiple sample records can |
| refer to different segments of one source video. The release preserves complete |
| source MP4 files and complete camera histories; the valid index selects target |
| chunks, rather than trimming away history required by the training sampler. |
|
|
| The dataset includes the existing video captions. Precomputed latent features |
| and depth maps are not included. |
|
|
| ## Layout |
|
|
| ```text |
| archives/videos-00000.tar # bounded video shards |
| archives/cameras-00000.tar.gz # portable camera annotations |
| metadata/samples.jsonl # sample-to-video mapping, existing caption, crop/cut |
| metadata/videos.jsonl # unique videos, byte sizes and SHA-256 |
| metadata/cameras.jsonl # camera record sizes and SHA-256 |
| metadata/valid_chunk_indices.json # unchanged selected feature-name -> chunk IDs |
| metadata/shards.json # archive sizes, members and SHA-256 |
| scripts/download_dataset.py # authenticated download, verification and extraction |
| release.json # dataset version and provenance |
| ``` |
|
|
| Extract all archives into the same dataset directory. Video members live under |
| `videos/` and annotation members under `cameras/`. Public paths are relative and |
| do not depend on internal storage mounts. |
|
|
| ## Camera and sampling conventions |
|
|
| Camera poses and intrinsics are estimated using **[Depth Anything 3](https://huggingface.co/depth-anything/DA3NESTED-GIANT-LARGE-1.1)** |
| (Nested Giant–Large, version 1.1), a model for reconstructing scene geometry from multiple images. |
| We sample 9 anchor frames from each 33-frame processed chunk at offsets |
| `[0, 4, 8, 12, 16, 20, 24, 28, 32]` and estimate them jointly with the video's reference frame. |
| The predicted poses are converted to camera-to-world transforms and expressed relative |
| to the video reference camera and to each chunk's first anchor. These annotations |
| are model estimates at the sampled anchors, rather than measured ground truth or dense per-frame poses. |
|
|
| - `camera_pose_rel_video_c2w`: `[chunks, 9, 4, 4]`, camera-to-world poses relative to the video's reference camera. |
| - `camera_pose_rel_chunk_c2w`: `[chunks, 9, 4, 4]`, poses relative to the first anchor of each chunk. |
| - `camera_K_norm`: `[chunks, 9, 4]`, normalized `[fx, fy, cx, cy]` using the model-processed image dimensions. |
| - `camera_valid_mask`: `[chunks, 9]`, validity flags. |
| - `camera_anchor_frame_idx`: source-video frame indices. |
| - `camera_anchor_processed_frame_idx`: indices after temporal selection. |
| - `selected_start_frame`, `selected_end_frame`, `stride`, `fps`, and the sample's original `cut` and `crop`: alignment metadata for reproducing preprocessing. |
|
|
| ## Download |
|
|
| Sign in to Hugging Face, complete the access form on this page, and accept the |
| agreement to receive download access automatically. Use your own account and |
| do not share access tokens: |
|
|
| ```bash |
| pip install huggingface_hub |
| hf auth login |
| hf download Drexubery/WorldCrafter --repo-type dataset --local-dir WorldCrafter-data |
| python WorldCrafter-data/scripts/download_dataset.py --root WorldCrafter-data --verify --extract |
| ``` |
|
|
| Keep approximately 400 GB free if retaining both downloaded archives and extracted |
| videos. Exact archive sizes and checksums are recorded in `metadata/shards.json`. |
|
|
