LottieGPT: A New Breakthrough in Enabling AI to Generate Vector Animations

The research team has achieved autoregressive generation of vector animations for the first time. Using a customized tokenizer and a dataset of 660,000 animations, they enabled multimodal models to directly generate editable Lottie animations, filling the gap in the field of vector animation generation.

In recent years, AI has made significant progress in video generation, but vector animation generation has long been overlooked. Vector animations have advantages such as resolution independence, compact file size, and editability. However, due to complex structures (layered geometric primitives, hierarchical relationships), time parameterization (keyframe interpolation), editability requirements, and sequence length challenges, existing models cannot synthesize them. Designers still need to create manually or rely on templates, which restricts creativity and large-scale applications.

Lottie (an open-source JSON format by Airbnb, cross-platform, semantically rich, and with a well-developed ecosystem) was chosen as the target format. A structure-aware Lottie Tokenizer was designed: geometric primitive encoding (Bezier curve parameterization), transformation matrix compression, keyframe representation (timestamp-attribute pairs), hierarchical structure preservation, and semantic alignment, reducing sequence length by an order of magnitude. LottieGPT was fine-tuned based on Qwen-VL, supporting text/image-to-animation generation and multimodal prompts, with output standard Lottie files that are editable.

The LottieAnimation-660K dataset was built: 660,000 real Lottie animations and 15 million static images, with diverse sources and quality control. Experimental results: The tokenizer reduces sequence length by 90% compared to the baseline while maintaining fidelity; generated animations score high in visual quality, coherence, and prompt adherence; generalize to different styles; outperform SOTA in SVG generation; outputs are editable by mainstream tools.

Application scenarios include: Designer assistant (rapid prototype generation), personalized content (non-professionals customizing animations), procedural generation (game/UI dynamic content), animation education (learning tools), and accessible design (visually impaired users creating via text).

Limitations: Quality of complex character animations and fine physical simulations needs improvement; insufficient coherence in long animations; lack of fine-grained control. Future directions: Hybrid diffusion and autoregressive architecture; interactive editing interface; expansion to formats like SVGA/Rive; exploration of video-to-vector animation conversion.