LottieGPT: Tokenizing Vector Animation for Autoregressive Generation

Why It Matters

This research addresses a significant gap in generative AI by enabling the creation of vector animations, a format crucial for web design, UI/UX, and motion graphics. The ability to generate editable, resolution-independent animations from natural language prompts has profound implications for design workflows, allowing for rapid prototyping and content generation.

Key Finding

AI can now generate editable vector animations by breaking them down into a token language, making it easier to create complex motion graphics from simple instructions.

Key Findings

• A tailored Lottie Tokenizer effectively encodes vector animation data into compact, semantically aligned token sequences.
• The LottieGPT model, trained on a large dataset, can generate coherent and editable vector animations from text or visual prompts.
• The proposed method significantly reduces sequence length while preserving structural fidelity, facilitating effective autoregressive learning.
• LottieGPT demonstrates strong generalization across diverse animation styles and outperforms existing SVG generation models.

Research Evidence

Aim: To develop a framework for tokenizing and autoregressively generating vector animations, specifically using the Lottie format, from natural language or visual prompts.

Method: Framework Development and Machine Learning

Procedure: The researchers developed a Lottie Tokenizer to encode layered geometric primitives, transforms, and keyframe-based motion into token sequences. They also curated a large dataset (LottieAnimation-660K) of Lottie animations and static images. This tokenizer and dataset were used to fine-tune a multimodal model (Qwen-VL) into LottieGPT, enabling it to generate vector animations from prompts.

Sample Size: 660,000 Lottie animations and 15 million static Lottie image files

Context: Digital Media Generation, AI-powered Design Tools

Design Principle

Leverage AI for structured data generation to create editable and resolution-independent digital assets.

How to Apply

Investigate existing AI models or research avenues that focus on generating structured vector graphics or animations for use in design projects.

Limitations

The effectiveness of the tokenizer and generation quality may depend on the diversity and quality of the training data; generalization to highly novel or complex animation styles might be limited.

Student Guide (IB Design Technology)

Simple Explanation: This research shows how computers can learn to create vector animations (like those used on websites) by turning them into a code-like language. This means AI could soon help designers make animations much faster.

Why This Matters: Understanding how AI can generate vector animations is important because it points to future tools that could significantly speed up the design process for motion graphics, UI elements, and interactive media.

Critical Thinking: To what extent can AI truly replicate the creative intent and nuanced artistry of a human animator, even with editable vector output?

IA-Ready Paragraph: Recent advancements, such as the LottieGPT framework, demonstrate the potential for AI to autoregressively generate editable vector animations by tokenizing animation data. This approach, which encodes geometric primitives and motion into token sequences, suggests future AI tools could significantly accelerate the creation of dynamic visual content for design projects.

Project Tips

• Consider how AI could automate repetitive animation tasks in your design project.
• Explore if existing AI tools can generate vector assets that you can then edit and refine.

How to Use in IA

• Reference this research when discussing the potential of AI in automating or assisting with the creation of visual assets, particularly animations.
• Use it to support claims about the future of generative design tools in your design project.

Examiner Tips

• When discussing AI in design, ensure you can explain *how* it works, not just that it exists. This paper provides a good example of a mechanism (tokenization) for AI-driven vector animation.
• Consider the implications of AI-generated editable assets on traditional design workflows.

Independent Variable: Natural language or visual prompts, Lottie animation data structure.

Dependent Variable: Coherence, editability, and quality of generated vector animations.

Controlled Variables: Lottie format specifications, underlying AI model architecture (Qwen-VL), tokenizer design.

Strengths

• Addresses a novel and important problem in generative AI (vector animation).
• Introduces a practical tokenization method and a large-scale dataset.
• Demonstrates strong performance on SVG generation.

Critical Questions

• How does the 'semantic alignment' of tokens impact the interpretability and editability of the generated animations?
• What are the ethical considerations of AI-generated animations, particularly regarding authorship and originality?

Extended Essay Application

• Investigate the potential for AI to generate interactive vector graphics for user interfaces, exploring user perception of AI-generated vs. human-designed elements.
• Develop a prototype tool that uses AI to generate animated infographics based on data inputs, evaluating its efficiency and effectiveness compared to manual methods.

Source

LottieGPT: Tokenizing Vector Animation for Autoregressive Generation · arXiv preprint · 2026