Effective integration of LLMs into animation design processes

Determine how to effectively integrate large language models into animation design processes to support animation ideation, creation, iteration, and refinement in ways that align with designers’ workflows and needs.

Background

The paper introduces Keyframer, a system that uses GPT-4 to generate CSS animations for SVG images from natural language prompts and supports iterative refinement through direct code and property editing. Through interviews and a user study, the authors document prompting strategies and the importance of tight feedback loops for animation design.

In the Discussion, the authors note that despite promising results, broader questions persist regarding how LLMs should be integrated into animation workflows. They outline several directions that highlight integration challenges, including model interpretability feedback, non-linear task decomposition, editing underlying graphics, combining prompting with direct manipulation, and designing for interactivity.

References

There remain several open questions about how LLMs might be effectively integrated into animation design processes.

— Keyframer: Empowering Animation Design using Large Language Models  (2402.06071 - Tseng et al., 2024) in Section 7: Discussion

Their pilot study reports that while the LLM can generate meaningful 3D call graphs for small programs that align with design principles, the quality of the generated output varies widely, leading them to raise open questions about whether LLMs can reliably function as visualization agents that construct 3D representations, reflecting cognitive design principles without explicit training.

— A Chat Assistant for Software Exploration in a 3D Software Visualization  (2610.09901 - Hansen et al., 7 Oct 2026) in Section 2, Related Work, paragraph “Tool-augmented LLM agents”