README.md

---
license: other
task_categories:
- image-to-3d
- image-to-image
tags:
- animals
- textures
- 3d
- meshes
- hair
- blender
- multimodal
pretty_name: AnimalLift
---

# AnimalLift

> **Official public dataset for the SIGGRAPH Asia 2026 paper**  
> **AnimalLift: Reconstructing Animatable 3D Animals from a Single Image via Canonical Shape, Texture, and Fur**

AnimalLift is the public dataset released with the SIGGRAPH Asia 2026 paper *AnimalLift: Reconstructing Animatable 3D Animals from a Single Image via Canonical Shape, Texture, and Fur*. The paper presents a framework for reconstructing animation-ready 3D animals from a single image using a shared canonical representation of shape, texture, and fur.

The dataset contains paired animal images, UV textures, shared-topology 3D meshes, and UV-aligned fur maps across multiple animal categories. These assets were generated to provide canonical supervision for geometry, appearance, and strand-level fur reconstruction.

The repository also includes `load_dataset_sample.py`, a Blender visualization script that loads one dataset sample, reconstructs hair curves, assigns materials, configures HDR environment lighting, and optionally saves a rendered image or `.blend` file. We test it on Blender 4.2.

## Dataset contents

The dataset is organized by animal category:

```text
AnimalLift/
├── dog/
│   └── dataset/
├── small_cat/
│   └── dataset/
├── big_cat/
│   └── dataset/
├── wolf/
│   └── dataset/
├── fox/
│   └── dataset/
├── bear/
│   └── dataset/
├── meta/
│   ├── hair_map_meta_uvanchor.npz
│   ├── render_elements/
│   │   └── hdr/
│   │       └── *.hdr
│   └── used_version_obj/
├── training_assets/
│   └── UV_Occupancy.png
└── load_dataset_sample.py
```

Each dataset group generally follows this structure:

```text
<animal>/dataset/<group>/
├── augmented_images/
│   └── *.png
├── textures/
│   └── *.png
├── shapes/
│   └── *.obj
├── hair_maps_single_512_uvlocal/
│   └── *.pt
└── hair_info.json
```

Some groups may also contain legacy `.npz` hair maps or optional deformed meshes.

## Modalities

### Augmented images

`augmented_images/` contains image inputs associated with the samples in a group.

### Textures

`textures/` contains UV texture maps. Texture files are paired with shapes using natural filename ordering.

### Shapes

`shapes/` contains OBJ meshes with UV coordinates.

### Hair maps

`hair_maps_single_512_uvlocal/` contains compressed UV-local hair representations.

The preferred `.pt` format stores:

```python
{
    "hair_q": ...,       # int8 tensor [H, W, S - 1, 3]
    "scale": ...,        # float32 tensor [3]
    "hair_mask": ...,    # bool tensor [H, W]
}
```

Legacy `.npz` files may contain:

```text
hair_offsets_local_q
offset_scale
samples
```

The global metadata file `meta/hair_map_meta_uvanchor.npz` provides the UV-to-surface mapping required to reconstruct strands:

```text
group_names
guide_group_id_map
uv_face_index_map
uv_bary_map
```

## Download

Clone the dataset repository:

```bash
git lfs install
git clone https://huggingface.co/datasets/Chunyi99/AnimalLift
cd AnimalLift
```

For large downloads, the Hugging Face CLI can also be used:

```bash
pip install -U huggingface_hub
hf download Chunyi99/AnimalLift \
  --repo-type dataset \
  --local-dir /path/to/animallift
```

## Blender visualization

The visualization entry point is:

```text
load_dataset_sample.py
```

### Requirements

- Blender with Python scripting support
- NumPy
- PyTorch in Blender's Python environment when loading `.pt` hair maps
- Blender procedural hair Geometry Nodes assets
- Cycles support for final rendering

The script uses Blender modules such as `bpy`, `bmesh`, and `mathutils`, so it must be executed by Blender rather than by a standard Python interpreter.

### Configure the dataset root

Open `load_dataset_sample.py` and set:

```python
ANIMALLIFT_ROOT = Path("/path/to/animallift")
```

The directory must contain the animal category folders and the `meta/` directory.

Do not commit a machine-specific absolute path.

### Run the default sample

```bash
blender --python load_dataset_sample.py
```

When arguments are supplied, place them after `--`:

```bash
blender --python load_dataset_sample.py -- \
  --species dog \
  --group_name groupA_different_breeds_chaiquan \
  --sample_index 0
```

### Load a specific OBJ

```bash
blender --python load_dataset_sample.py -- \
  --obj_path /path/to/animallift/dog/dataset/groupA_different_breeds_chaiquan/shapes/aug_01.obj
```

The script resolves the texture and hair map from the same dataset group.

### Save a Blender scene

```bash
blender --python load_dataset_sample.py
```

## Command-line arguments

| Argument | Description |
|---|---|
| `--obj_path` | Optional path to a specific OBJ inside a group `shapes/` directory. |
| `--species` | Animal category used for automatic sample discovery. |
| `--group_name` | Dataset group used for automatic sample discovery. |
| `--sample_index` | Natural-sort index of the sample inside the selected group. |
| `--mesh_object_name` | Name assigned to the imported Blender mesh object. |
| `--deformed_obj_path` | Optional deformed OBJ path. |
| `--shape_key_name` | Name used when applying deformation as a shape key. |
| `--shape_key_value` | Shape-key influence value. |
| `--save_blend_path` | Optional output path for a `.blend` file. |
| `--render_filename` | Optional output render filename. |
| `--use_shape_key` | Use the legacy shape-key deformation path. |


## Paper

This dataset accompanies the SIGGRAPH Asia 2026 paper:

**AnimalLift: Reconstructing Animatable 3D Animals from a Single Image via Canonical Shape, Texture, and Fur**

The paper introduces:

- a canonical shared-topology representation for animal geometry, texture, and fur;
- a procedural dataset generation pipeline covering diverse species, body shapes, poses, textures, and fur properties;
- a UV-aligned fur map that explicitly encodes strand geometry and supports deformation, animation, and simulation;
- a single-image feed-forward reconstruction framework that jointly predicts canonical shape, texture, and fur.

The current manuscript describes a dataset with shared topology and UV parameterization, including 3,000 dog shapes, 3,000 cat shapes, 500 lion and tiger shapes, and 1,500 shapes each for wolves, foxes, and bears. Textures are generated at 1024 × 1024 resolution, while the canonical fur map is defined on a 512 × 512 UV grid.

## Citation

The uploaded manuscript is currently anonymized and contains a placeholder DOI. Until the final bibliographic record is available, cite the project using the temporary entry below:

```bibtex
@article{animallift2026,
  title   = {AnimalLift: Reconstructing Animatable 3D Animals from a Single Image via Canonical Shape, Texture, and Fur},
  author  = {Anonymous Authors},
  year    = {2026},
  note    = {SIGGRAPH Asia 2026}
}
```

Replace the author list, volume, issue, article number, pages, and DOI with the final publication metadata once released.


## Indexed content

- Samples: 11545
- Uploaded data files: 48490
- Data size: 295.51 GiB
- Guidance directory: `augmented_images`
- Texture directory: `textures`
- Shape directory: `shapes`
- Hair directory: `hair_maps_single_512_uvlocal`
- Recursive indexing: `False`

Each species is stored under:

```text
<species>/dataset/<original relative path>
```

Original group names are replaced by deterministic names such as `bear000001`.
Numbering restarts within each species. The upload uses hard links, so no full
data copy is created. `manifest.json` records the original-to-new name mapping.