MoVer: A tool that creates animations via LLM-based iterative refinement (mover-dsl.github.io)

MoVer is an innovative tool designed to improve the creation of motion graphics animations by combining large language models (LLMs) with a specialized verification framework. Traditional vision-language models often struggle to fully capture the spatio-temporal details—such as direction, timing, and relative positioning—described in text prompts when generating animations. To tackle this, MoVer introduces a domain-specific language (DSL) based on first-order logic that verifies these motion properties within SVG-based animations, enabling precise checks against the intended animation behavior. What makes MoVer particularly significant is its integration into an iterative synthesis and verification pipeline: the LLM first generates an animation and a corresponding MoVer verification program; the program then identifies which motion predicates fail, and this feedback is automatically used to refine the animation iteratively. This loop substantially boosts accuracy—while the pipeline achieves correct animations on 58.8% of prompts without iteration, it reaches 93.6% correctness after up to 50 refinement cycles. The team validated MoVer on a robust synthetic dataset of 5,600 text prompts, demonstrating how formal verification combined with LLM reasoning can address the common challenge of incomplete or inconsistent motion generation. Technically, MoVer’s verification engine supports a general set of spatio-temporal predicates expressed in a logic-based DSL, allowing precise auditing of animation properties across frames. This approach not only enhances the reliability of AI-driven motion design but also establishes a new paradigm for incorporating formal specifications into creative AI workflows, promising wider application in visual programming and automated animation generation.