An interactive evaluation and preview panel I built for Code-to-Video agents. It is a standalone page.
Open Round 2 main panel → · Open Round 1 pilot panel →
In this evaluation, coding agents take a short open-ended prompt (with optional reference assets), write rendering code (HTML + headless browser, Canvas/WebGL, Three.js, frame-by-frame Python, or HyperFrames), and deliver an MP4 via ffmpeg. It covers:
- 15 tasks across two rounds with 136 rendered videos and full tool-call trajectories, supporting per-task overviews, side-by-side comparisons (
#/compare), frame strips, and the frames inspected by the agent itself (seen/) - Six models: Claude Opus 5.5, GPT-6 Astra, DeepSeek V4.1-Flash, GPT-5.6 Sol, plus generational control Claude Opus 5 and text-only control DeepSeek V4-Pro, tested under bare prompts (
bare) and pre-initialized HyperFrames workflows with skills (hf) - Round 2 tasks spanning science explanation (
S1), wordless geometry proof (M1), first-person robotic grasping (E1), painting animation (I1), music videos (B1/B2), three-act paper-plane storytelling (C1), Titanic bow recreation (R1), four-camera needle threading (P1), three-camera soft-body pinching (P2), and reference video recreation (V2) - Automated judging with Gemini 4 Argon (frozen binary checklists, swap-consistent blind pairwise comparisons, reverse/blur perturbation controls, and an independent
subtextdimension) alongside physical, narrative, and audio-visual alignment analyses
Text reports (Chinese): synthesis report · capabilities & skill internalization · frontier directions & rubric design · physical & 3D highlights · narrative & music highlights · Argon v2 report · Argon v3 subtext report