Gemini 1.5 Flash

The native language of video synthesis When we force large language models to output custom JSON schemas or learn specialized video editing domain-specific languages, we create unnecessary friction. They perform poorly because we force them to speak a foreign tongue. Instead, the team at HeyGen embraced a simpler truth: large language models already master HTML, CSS, and JavaScript. Web code makes up the vast majority of web-scraping training data. By building on top of this native foundation, the open-source tool Hyperframes lets AI agents output video layouts using the exact code they already generate best. Why Remotion fell short for AI agents Many developers look to React-based frameworks like Remotion for programmatic video editing. While excellent for human developers, these frameworks restrict AI creativity. The structure of React boxes the model in, requiring extensive instruction and highly complex prompt engineering just to yield basic, safe results. Adding guardrails often sterilizes the creative output. To avoid this, Hyperframes uses a incredibly thin wrapper around plain HTML with a few custom data attributes. During development, the team utilized Gemini 1.5 Flash as a design partner. The philosophy was simple: if a small, efficient model can easily author the code, advanced agents will execute it flawlessly. Solving the asynchronous browser render trap Browsers are built to load assets asynchronously to keep web surfing fast. Fonts, images, and scripts load whenever they are ready. While this makes websites feel smooth, it is disastrous for deterministic video production where every single frame must look identical on every render. To solve this, Hyperframes freezes the browser's internal clock. The rendering engine steps through the timeline frame by frame, waiting explicitly for every asset to fully load before snapping a high-resolution screenshot. These individual screenshots are then compiled into a pixel-perfect MP4. This process unlocks advanced native web technologies like Three.js and WebGL directly inside video workflows. Raising the floor of AI video generation Instead of teaching agents how to code from scratch, developers can focus on teaching them taste. By pairing the framework with specific video skills, the system acts as a creative director rather than a syntax compiler. This keeps humans in the loop for final polishes, allowing automated agents to draft launch videos, interactive product tours, and high-fidelity animations using the same codebase as the applications they showcase.