OmniLottie: Decoding the Future of Intelligent Vector Animation

TL;DR

OmniLottie is the first end-to-end framework capable of generating professional-grade Lottie vector animations from text, images, or videos. By reinventing how Lottie files are "read" by AI through a specialized tokenizer and providing a massive new dataset (MMLottie-2M), it bridges the gap between static design and dynamic motion, achieving a staggering 97.3% success rate where traditional LLMs often fail.

The "JSON Tax": Why Vector Animation is Hard for AI

Vector animation, particularly the Lottie format, is a gold standard for web and mobile UI because it's resolution-independent and lightweight. However, for a Large Language Model (LLM), a Lottie file is a nightmare. A typical Lottie JSON is filled with "boilerplate" metadata—indentation, curly braces, and redundant structural tags—that account for over 80% of the file size.

When we ask a model like GPT-4 or Qwen to generate Lottie code directly, the model wastes its "attention budget" on reproducing syntax rather than understanding geometry or physics. This results in broken files, collapsed shapes, or simply "hallucinated" code that won't render.

Methodology: The Lottie Tokenizer

The core breakthrough of OmniLottie is moving from Text-to-JSON to Text-to-Command.

1. Stripping the Redundancy

Instead of teaching the model to write JSON, the authors developed a Lottie Tokenizer. This tool parses the complex hierarchy of a Lottie file into a flat sequence of Animation Commands.

• Geometric Transforms: Position, Scale, Rotation.
• Visual Properties: Fill colors, Stroke widths.
• Temporal Cues: Keyframes, In-points, Out-points.

2. Offset-Based Parameterization

To keep the vocabulary clean, they use an offset-based scheme. Different "types" of values (like a coordinate vs. an opacity percentage) are mapped to different regions of the token space. This prevents the model from confusing a "red color code" with a "50-pixel movement."

Figure 1: The OmniLottie Pipeline – From Multi-modal instructions to Parameterized Token Sequences.

MMLottie-2M: A Foundation for Motion

Research in this field was previously stifled by a lack of data. OmniLottie introduces MMLottie-2M, a dataset of 2 million Lottie animations.

• Hybrid Sourcing: It combines professionally designed animations from the web with synthetic animations created by applying motion templates to static SVGs.
• Multi-modal pairing: Each animation is paired with rich text descriptions, keyframe images, and reference videos, allowing the model to learn the relationship between "what it is" and "how it moves."

Performance: Success vs. Failure

The results are transformative. Previous SOTA models like DeepSeek-V3 or GPT-5 struggle significantly with the strict formatting of Lottie, often achieving success rates below 10-30%. OmniLottie, thanks to its structured vocabulary, hits 88-97%.

Figure 2: Quantitative comparison across different tasks. OmniLottie (Ours) consistently leads in success rate and visual fidelity (FVD).

Why it wins:

1. Speed: It is 530x faster than optimization-based methods (which try to "evolve" a vector until it matches a video).
2. Editability: Unlike Video-Generation models (which produce pixels), OmniLottie produces pure vector code that you can open in Adobe After Effects and tweak.
3. Semantic Control: It accurately follows complex instructions like "A blue paper plane flies from left to right" while maintaining the stylistic integrity of the input image.

Critical Analysis & Future Outlook

OmniLottie proves that structured representation beats raw generation. By treating Lottie files as a sequence of function calls rather than a text file, the authors unlocked the potential of VLMs in the design space.

Limitations:

• Complexity Bloom: For extremely complex animations with hundreds of layers, the token sequence length might still exceed the model's context window.
• Invalid Sequences: While rare, the model can still generate "rendering failures" (e.g., a shape with no fill color).

The Future: We are moving toward a world where designers use "Agentic Design Tools." Instead of manually setting keyframes, a developer might simply say, "Make this logo pulse like a heartbeat whenever a user hovers over it," and OmniLottie will generate the precise Lottie payload instantly.