PAN 2: simulating what happens next, before making the scene
Analysis of IFM’s November 11, 2026 announcement, not a new launch today. PAN 2 connects an initial scene and a written action to a simulated video.
MBZUAI / Robotics lab — press-use photograph from https://mbzuai.ac.ae/news-events/media-center, used with attribution; converted to WebP. This shows a university lab, not PAN 2 output or an independent model test. / جامعة محمد بن زايد للذكاء الاصطناعي، مختبر الروبوتات — صورة مخصصة للاستخدام الصحفي مع نسبها إلى الجامعة، حُوّلت إلى WebP. لا تعرض نتيجة من PAN 2 أو اختبارًا مستقلًا للنموذج.
According to MBZUAI, the Institute of Foundation Models combines language-based planning with video diffusion in PAN 2. IFM says it released video-model weights, training and inference code, and documentation under Apache 2.0. The announcement also identifies long-term world consistency as an open challenge.
For creative work, a useful test would be to describe one action and compare the simulated motion with the intended scene: does an object persist, does contact make sense, and does the camera stay coherent? These are proposed evaluation questions, not results from a test performed here. A plausible video should not be treated as proof of physical accuracy.