Sssaasss/MMLottie-2M

---
license: cc-by-nc-sa-4.0
language:
- en
tags:
- lottie
- animation
- vector-graphics
- motion-graphics
- multi-modal
size_categories:
- 1M<n<10M
configs:
- config_name: Lottie
data_files: data/Lottie/*.parquet
- config_name: Lottie_SVG
data_files: data/Lottie_SVG/*.parquet
---

# MMLottie-2M Dataset

The first large-scale Lottie animation dataset for multi-modal vector animation generation, containing ~2M samples with diverse motion patterns and visual styles.

## Dataset Overview

**MMLottie-2M** consists of two complementary subsets designed to support comprehensive training for Lottie animation generation:

### 1. Lottie Subset

**Native Lottie animations** collected from major online platforms including LottieFiles, IconScout, Flaticon, Iconfont, and Icons8.

**Data Processing:**
- Removal of irrelevant elements (base64 images, non-visual layers, After Effects expressions)
- Filtering of non-parameterizable layers
- Spatial normalization to 512×512 canvas
- Temporal normalization to 0-16 timestamp range
- Center alignment with aspect ratio preservation

**Purpose:** Provides authentic motion graphics with complex layer structures and real-world motion patterns.

### 2. Lottie_SVG Subset

**SVG-to-Lottie converted animations** generated from the large-scale OmniSVG collection with motion augmentation.

**Generation Process:**
- Base: Static SVG files from MMSVG-2M dataset
- Motion Transfer: 1,678 canonical motion templates extracted from native Lottie files
- Motion Patterns: Translations, zooms, rotations, opacity changes, and combinations
- Augmentation: Automated keyframe injection to create diverse motion dynamics

**Purpose:** Decouples visual content from motion semantics, enabling better alignment between visual components and animation conditions. Reduces the path distribution gap and increases animated layer coverage for improved model training.

**Key Characteristics:**
- Motion signatures encoding temporal patterns (e.g., "fade-in + upward motion + scale-down")
- Semantically clustered motion templates with caption keywords
- Reduces path distribution gap from 24% to <1%
- Increases animated layer coverage from 0% to 16%

## Usage

### Load specific configuration

```python
from datasets import load_dataset

# Load native Lottie animations
dataset_lottie = load_dataset("OmniLottie/MMLottie-2M", "Lottie")

# Load SVG-based Lottie animations with motion augmentation
dataset_svg = load_dataset("OmniLottie/MMLottie-2M", "Lottie_SVG")
```

### Load subset of data

```python
# Load first 1000 samples from Lottie_SVG
dataset_subset = load_dataset("OmniLottie/MMLottie-2M", "Lottie_SVG", split="train[:1000]")

# Load 10% of Lottie data
dataset_10pct = load_dataset("OmniLottie/MMLottie-2M", "Lottie", split="train[:10%]")
```

### Load all configurations

```python
# Load both configurations together
dataset_all = load_dataset("OmniLottie/MMLottie-2M")
```

## Dataset Fields

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique identifier (MD5 hash) |
| `source` | string | Data source ("Lottie" or "Lottie_SVG") |
| `lottie_json` | string | Normalized Lottie JSON (512×512, 0-16 frames) |
| `image` | Image | PNG preview image |
| `video` | Video | MP4 animation (h264 encoding, random light background) |
| `detail` | string | Detailed caption (subjects, objects, motion, color, style) |
| `desc_en` | string | English description with temporal details |
| `keywords_en` | string | Keywords emphasizing geometry and motion |
| `token_length` | int64 | Token length of Lottie JSON |
| `motion_type` | string | Motion pattern type (Lottie_SVG only) |
| `motion_caption` | string | Motion-specific caption (Lottie_SVG only) |

## Supported Tasks

This dataset supports three multi-modal vector animation generation tasks:

1. **Text-to-Lottie**: Generate Lottie animations from text descriptions
2. **Image-Text-to-Lottie**: Generate animations from image + text (foreground motion focus)
3. **Video-to-Lottie**: Generate parameterized Lottie from video demonstrations

## Data Annotation

Annotations are generated using Vision-Language Models (VLMs) with a coarse-to-fine strategy:

1. **Coarse**: Overall caption covering subjects, objects, motion, color, and style
2. **Fine**: Temporal details across frames with cues like "begins with" and "then"
3. **Emphasis**: Keywords highlighting geometry and motion for better text-following

## Citation

If you use this dataset, please cite:

```bibtex
@article{yang2026omnilottie,
title={OmniLottie: Generating Vector Animations via Parameterized Lottie Tokens},
author={Yiying Yang and Wei Cheng and Sijin Chen and Honghao Fu and Xianfang Zeng and Yujun Cai and Gang Yu and Xinjun Ma},
journal={arXiv preprint arxiv:2603.02138},
year={2026}
}
```

## Acknowledgments

We thank the following projects and resources for their valuable contributions:

- **Data Sources**: [LottieFiles](https://lottiefiles.com), [IconScout](https://iconscout.com), [Flaticon](https://www.flaticon.com), [Iconfont](https://www.iconfont.cn), [Icons8](https://icons8.com)
- **[python-lottie](https://github.com/eltiempoes/python-lottie)**: For providing excellent tools for Lottie manipulation and processing
- **[MMSVG-Icon](https://huggingface.co/datasets/OmniSVG/MMSVG-Icon)**, **[MMSVG-Illustration](https://huggingface.co/datasets/OmniSVG/MMSVG-Illustration)**: For inspiring our multi-modal data curation approach