Can language models make better videos than video models?
The author compares the video generation capabilities of language models and dedicated video models by testing GPT-6.2 Astra and Dreamina Seedance 2.5 on identical 3D scenes. GPT-6.2 Astra generated the scenes using three.js code and motion graphics, while Dreamina Seedance 2.5 used text prompts and reference images. Both models were tested with a scripted 15-second camera path at 1080p and 24 fps. Open article →