AnimalLift

Official dataset for AnimalLift · SIGGRAPH Asia 2026

AnimalLift: Reconstructing Animatable 3D Animals from a Single Image by Learning Canonical Shape, Texture, and Fur Maps

Chunyi Sun¹ · Ruyi Zha¹ · Weijian Deng¹ · Junlin Han² · Dylan Campbell¹ · Stephen Gould¹

¹ Australian National University   ² University of Oxford

Code · Model & Checkpoints · Dataset Files

Overview · Download · Dataset Structure · Blender Visualization · Citation · License

Overview

AnimalLift reconstructs animation-ready 3D animals from a single image using a shared canonical representation of shape, texture, and fur. This dataset provides paired animal images, UV textures, shared-topology meshes, and UV-aligned fur maps for supervising geometry, appearance, and strand-level fur reconstruction.

Property	Released data
Indexed samples	11,545
Uploaded data files	48,490
Data size	295.51 GiB
Animal categories	Dog, small cat, big cat, wolf, fox, bear
Texture resolution	1024 × 1024
Canonical fur-map resolution	512 × 512
Tested visualization environment	Blender 4.2

The dataset includes load_dataset_sample.py to visualize an individual sample in Blender. It imports a mesh, reconstructs hair curves, assigns materials, configures HDR environment lighting, and optionally exports a rendered image or a .blend scene.

For model installation, training, and single-image inference, see the code repository. Pretrained weights are available in the model repository.

Download

Hugging Face CLI

Download the complete dataset to a directory of your choice:

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

Allow sufficient storage for the approximately 295.51 GiB download.

Git LFS

Alternatively, clone the dataset repository with Git LFS installed:

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

Dataset Structure

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
├── manifest.json
└── load_dataset_sample.py

Each group generally has the following structure:

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

Published group names are deterministic identifiers such as bear000001, with numbering restarting for each species. manifest.json records the mapping between original and published names. Use the published folder names when selecting a group in Blender.

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

Modalities

Directory	Contents
augmented_images/	Image inputs associated with the group's samples.
textures/	UV texture maps, paired with shapes using natural filename ordering.
shapes/	OBJ meshes with UV coordinates and shared topology.
hair_maps_single_512_uvlocal/	Compressed UV-local strand representations.

Hair-map format

The preferred .pt format stores:

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

Here, H and W are the UV-grid dimensions and S is the number of samples along a strand.

Legacy .npz files may contain hair_offsets_local_q, offset_scale, and samples.

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

group_names
guide_group_id_map
uv_face_index_map
uv_bary_map

Blender Visualization

Requirements

The dataset visualization script has been tested with Blender 4.2. It requires:

• NumPy in Blender's Python environment.
• PyTorch in Blender's Python environment to read .pt hair maps.
• Blender procedural hair Geometry Nodes assets.
• Cycles support when producing a final render.

Run the script with Blender. It imports bpy, bmesh, and mathutils, which are Blender modules. Installing PyTorch only in a separate Conda environment does not make it available to Blender's Python.

1. Configure the dataset root

Set ANIMALLIFT_ROOT in load_dataset_sample.py to your downloaded dataset directory:

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

This directory must contain the animal category folders and meta/. Run the following examples from that directory.

2. Open a sample

Launch the script with its default sample selection:

blender --python load_dataset_sample.py

To select a group, replace <group> with an existing folder name under dog/dataset/. Arguments for the script go after --:

blender --python load_dataset_sample.py -- \
  --species dog \
  --group_name "<group>" \
  --sample_index 0

To load a specific OBJ, replace the path below with an existing mesh path:

blender --python load_dataset_sample.py -- \
  --obj_path "/path/to/animallift/dog/dataset/<group>/shapes/<sample>.obj"

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

3. Save a .blend scene on a server

Use background mode to create a scene without opening the Blender interface:

mkdir -p outputs
blender --background --python-exit-code 1 \
  --python load_dataset_sample.py -- \
  --save_blend_path outputs/sample.blend

Add --species, --group_name, and --sample_index, or --obj_path, to select a particular sample. Open the saved .blend file in Blender 4.2 to inspect the scene interactively.

Command-line reference

Argument	Description
--obj_path	Path to a specific OBJ inside a group's shapes/ directory.
--species	Animal category for automatic sample discovery.
--group_name	Published dataset group name for automatic sample discovery.
--sample_index	Natural-sort index of the sample in the selected group.
--mesh_object_name	Name assigned to the imported mesh object.
--deformed_obj_path	Optional deformed OBJ path.
--shape_key_name	Name used when applying a deformation as a shape key.
--shape_key_value	Shape-key influence value.
--save_blend_path	Output path for a .blend scene.
--render_filename	Optional output render filename.
--use_shape_key	Enable the legacy shape-key deformation path.

Citation

If you use AnimalLift in your research, please cite:

@inproceedings{sun2026animallift,
  title     = {AnimalLift: Reconstructing Animatable 3D Animals from a Single Image by Learning Canonical Shape, Texture, and Fur Maps},
  author    = {Sun, Chunyi and Zha, Ruyi and Deng, Weijian and Han, Junlin and Campbell, Dylan and Gould, Stephen},
  booktitle = {SIGGRAPH Asia 2026 Conference Papers},
  year      = {2026}
}

License

The AnimalLift dataset is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0).

You may copy and redistribute the dataset for noncommercial purposes under the license terms. You must give appropriate credit, retain the required notices, and provide a link to the license. If you remix, transform, or build upon the dataset, you may not distribute the modified material.

See the full license text for the complete terms.