LottieGPT/LottieSVG-10M

LottieGPT

Tokenizing Vector Animation for Autoregressive Generation

Junhao Chen^1*, Kejun Gao^1*, Yuehan Cui¹, Mingze Sun¹, Mingjin Chen³
Shaohui Wang¹, Xiaoxiao Long⁴, Fei Ma⁵, Qi Tian⁵, Ruqi Huang^1†, Hao Zhao^1,2†

¹ Tsinghua University ² BAAI
³ The Hong Kong Polytechnic University
⁴ Nanjing University ⁵ Guangming Lab

---

Overview

LottieGPT is a model for generating editable vector animations in an autoregressive manner. Instead of producing fixed-resolution raster frames, it tokenizes vector animation structure and motion, enabling high-quality generation that remains editable after synthesis.

Highlights

• LottieSVG-10M: a large-scale SVG-to-Lottie dataset with rendered PNG images, text captions, SVG code, and converted Lottie JSON
• LottieAnimation-660K: a large-scale Lottie animation dataset with MP4 previews, text captions, tags, and full Lottie JSON
• Editable outputs: generated animations can be directly edited at the shape and motion level
• Autoregressive tokenization: vector animation is modeled as a learnable token sequence
• Multi-modal conditioning: supports text, image, and keyframe-based generation

Open-source Plan

• Project Page & Technical Report
• LottieSVG-10M & LottieAnimation-660K Dataset Release
• Inference Code & Model Weight
• Online Demo
• LottieBench Benchmark
• Training Code

Paper Links

• Paper: LottieGPT: Tokenizing Vector Animation for Autoregressive Generation
• CVPR Open Access: https://openaccess.thecvf.com/content/CVPR2026/html/Chen_LottieGPT_Tokenizing_Vector_Animation_for_Autoregressive_Generation_CVPR_2026_paper.html
• arXiv: https://arxiv.org/abs/2604.11792
• PDF: https://arxiv.org/pdf/2604.11792
• Project Page: https://lottiegpt.github.io/
• GitHub: https://github.com/yisuanwang/LottieGPT

LottieSVG-10M

This release contains the LottieSVG-10M dataset introduced with LottieGPT. It provides SVG code, converted Lottie JSON, rendered PNG images, tags, and captions for large-scale vector graphics generation research.

Dataset Size

• Split: train
• Examples: 9,778,526
• Metadata shards: 100 compressed JSONL files, metadata-xxxxx-of-00100.jsonl.zst
• Image shards: 100 compressed tar files, images-xxxxx-of-00100.tar.zst
• Removed from release: 17,291 source candidates without valid lottie_json

License

This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0) for academic research and non-commercial use only.

Commercial use, redistribution for profit, or deployment in commercial products or services is not permitted without explicit written authorization from the dataset authors. The license applies to the dataset organization, metadata, annotations, and processed release files provided here; it does not override any rights held by the original content owners.

Directory Layout

LottieSVG-10M/
  README.md
  dataset_info.json
  stats.json
  validation_summary.json
  load_dataset.py
  data/
    metadata-00000-of-00100.jsonl.zst
    ...
    images-00000-of-00100.tar.zst
    ...

Metadata Schema

Each JSONL row contains:

{
  "id": "6476775",
  "caption": "A gradient blue-to-purple light bulb is surrounded by radiating purple lines, representing a glowing effect.",
  "tag": ["idea", "creative", "bulb"],
  "image": {
    "tar": "images-00000-of-00100.tar",
    "key": "00000/00000000_6476775.png",
    "format": "png"
  },
  "svg_code": "<svg ...></svg>",
  "lottie_json": {"v": "..."},
  "asset": "icon",
  "width": 512,
  "height": 512,
  "svg_filename": "Idea_original.svg",
  "style": "Gradient"
}

How to Use This Dataset

The release is organized for large-scale training without creating millions of small files. Metadata is stored as compressed JSONL shards, and rendered PNG images are packed into compressed tar shards.

1. Install dependencies

pip install zstandard pillow
# Optional command-line decompressor on Ubuntu/Debian:
# sudo apt-get install zstd

The Python examples below need zstandard for .zst metadata and Pillow only when opening PNG previews.

2. Stream metadata directly from .jsonl.zst

You do not need to decompress metadata to disk. Each line is one training example.

import io
import json
from pathlib import Path
import zstandard as zstd

root = Path("/path/to/LottieSVG-10M")
meta_path = root / "data" / "metadata-00000-of-00100.jsonl.zst"

with meta_path.open("rb") as fh:
    reader = zstd.ZstdDecompressor().stream_reader(fh)
    text = io.TextIOWrapper(reader, encoding="utf-8")
    for line in text:
        row = json.loads(line)
        print(row["id"], row["caption"])
        print(row["svg_code"][:80])
        print(row["lottie_json"].keys())
        break

Important fields:

• id: unique sample id
• caption: text caption
• tag: keyword list
• svg_code: original SVG code
• lottie_json: converted Lottie JSON object
• image.tar: tar shard containing the rendered PNG
• image.key: path of the PNG inside the tar shard

3. Decompress one image tar shard before reading PNGs

The metadata image reference uses the uncompressed tar name, for example images-00000-of-00100.tar. In this release the stored file is images-00000-of-00100.tar.zst. Decompress the needed shard first:

zstd -d -k data/images-00000-of-00100.tar.zst

The -k flag keeps the compressed .tar.zst file. After decompression, the directory should contain both:

data/images-00000-of-00100.tar.zst
data/images-00000-of-00100.tar

Then read the PNG by image.key from the uncompressed tar:

import io
import json
import tarfile
from pathlib import Path
from PIL import Image
import zstandard as zstd

root = Path("/path/to/LottieSVG-10M")
meta_path = root / "data" / "metadata-00000-of-00100.jsonl.zst"
with meta_path.open("rb") as fh:
    text = io.TextIOWrapper(zstd.ZstdDecompressor().stream_reader(fh), encoding="utf-8")
    row = json.loads(next(text))

image_ref = row["image"]
with tarfile.open(root / "data" / image_ref["tar"], "r") as tar:
    image_bytes = tar.extractfile(image_ref["key"]).read()
image = Image.open(io.BytesIO(image_bytes))
print(image.size)

4. Decompress all image shards for repeated random access

For one-off metadata training, keep the files compressed and stream the metadata. For repeated image access, decompress the image tar shards once:

find data -name 'images-*.tar.zst' -print0 | xargs -0 -n 1 -P 8 zstd -d -k -f

Use a smaller or larger -P value depending on your CPU and disk bandwidth. Keep the .zst files if you plan to redistribute the compressed release.

5. Use the convenience loader

load_dataset.py can iterate compressed metadata directly. To include image bytes, first decompress the referenced image tar shards.

python load_dataset.py /path/to/LottieSVG-10M --limit 3
python load_dataset.py /path/to/LottieSVG-10M --limit 3 --include-image-bytes

You can also import the iterator in training code:

from load_dataset import iter_lottiesvg

for row in iter_lottiesvg("/path/to/LottieSVG-10M"):
    svg_code = row["svg_code"]
    lottie_json = row["lottie_json"]
    caption = row["caption"]
    tags = row["tag"]
    break

To read PNG bytes through the loader, decompress the referenced image tar shards first and pass include_image_bytes=True.

Validation

Release validation checks that every retained row has non-empty id, caption, svg_code, embedded lottie_json, and a PNG reference under image.tar/image.key. The rendered images are stored in tar shards to avoid millions of small files.

Dataset Disclaimer

Intended Use

The LottieSVG-10M dataset is provided exclusively for academic research and non-commercial purposes. Any commercial use, redistribution for profit, or deployment in commercial products or services is prohibited without explicit written authorization from the dataset authors.

Data Source and Intellectual Property

The dataset is compiled, processed, filtered, captioned, rendered, converted, and reorganized from content that was originally publicly available from third-party sources. All copyrights, trademarks, and other intellectual property rights in the original content remain with their respective owners.

The inclusion of any content in this dataset does not imply endorsement, authorization, sponsorship, or affiliation with the original creators, platforms, or rights holders. The processing performed by the authors, including filtering, SVG-to-Lottie conversion, rendering, captioning, and packaging, does not alter the ownership or intellectual property status of the underlying content.

No Warranties

The dataset is provided "AS IS" and "AS AVAILABLE", without warranties of any kind, either express or implied, including but not limited to warranties of accuracy, completeness, reliability, merchantability, fitness for a particular purpose, non-infringement of third-party rights, or freedom from errors.

Limitation of Liability

Under no circumstances shall the authors, contributors, or affiliated organizations be liable for any direct, indirect, incidental, special, consequential, or punitive damages arising from or related to the use of, inability to use, or reliance on this dataset, including claims by third parties regarding intellectual property rights.

User Responsibilities

By using this dataset, you agree that you are solely responsible for ensuring compliance with all applicable laws, regulations, and third-party rights in your jurisdiction. You agree not to use the dataset for illegal, harmful, unethical, or commercial purposes, and to properly cite LottieGPT in any resulting publication or research output.

Content Removal Requests and Contact

If you are a rights holder and believe that any content in this dataset infringes your intellectual property rights, please contact us. We will review legitimate removal requests and address verified claims.

For questions, collaborations, permission inquiries, or content removal requests, please contact dreamhowchen@gmail.com.

LottieGPT Paper and Citation

If you use this dataset or LottieGPT in your work, please cite:

@InProceedings{Chen_2026_CVPR,
  author = {Chen, Junhao and Gao, Kejun and Cui, Yuehan and Sun, Mingze and Chen, Mingjin and Wang, Shaohui and Long, Xiaoxiao and Ma, Fei and Tian, Qi and Zhao, Hao and Huang, Ruqi},
  title = {LottieGPT: Tokenizing Vector Animation for Autoregressive Generation},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2026},
  pages = {31639-31651}
}

Project resources:

• Paper: https://openaccess.thecvf.com/content/CVPR2026/html/Chen_LottieGPT_Tokenizing_Vector_Animation_for_Autoregressive_Generation_CVPR_2026_paper.html
• Project page: https://lottiegpt.github.io/
• GitHub: https://github.com/yisuanwang/LottieGPT