Generating Motion Graphic Videos via LLM Code Rendering Instead of Video Diffusion
A Reddit user demonstrated a novel approach to video creation by prompting GLM 5.3 Flash to generate executable HTML5 Canvas animation code instead of relying on traditional AI video diffusion models. The setup, served using vLLM, produced a 40-second stock market infographic video guided by a hand-drawn canvas animation skill pack. This code-driven approach offers deterministic rendering, perfectly legible text, and sharp visual components without the visual artifacts or high inference costs of video diffusion models. It highlights how code-generation LLMs can be harnessed to automate programmatic motion graphics and visual content creation. The workflow relies on a custom skill pack defining hand-drawn Canvas graphics rules, allowing the model complete autonomous control to generate the animation code from a broad prompt. The GLM model was run in 8-bit precision using the vLLM inference engine on a four-DGX hardware setup.
## BACKGROUND
AI video generation traditionally relies on diffusion models that synthesize pixel frames directly, which often struggle with precise text, charts, or mathematical graphics. In contrast, code generation utilizes Large Language Models (LLMs) such as Zhipu AI's GLM family to output vector scripts or JavaScript code that render smoothly in a browser. High-throughput inference engines like vLLM are commonly used to serve these models efficiently.