Mango AI: image/video generation with Nano Banana 2, GPT Image 2, Seedance 2

Meta's codename Mango, a next-generation multimodal image and video generation AI model developed by Meta Superintelligence Labs (MSL) led by Alexandr Wang, was revealed during an internal Q&A session on December 18, 2025. Built on a diffusion-transformer (DiT) architecture, Mango learns physics alongside visual synthesis to reduce hallucinations in generated videos.

Mango is Meta's codename for a next-generation multimodal image and video generation AI model , developed inside Meta Superintelligence Labs MSL — Meta's elite AI research division led by Alexandr Wang. First revealed during an internal Q&A session on December 18, 2025, Mango represents Meta's most ambitious push into generative media. Built on a diffusion-transformer DiT architecture , Mango goes beyond simple pixel generation — it learns the laws of physics alongside visual synthesis, enabling generated objects to maintain realistic shape, mass, and velocity over time. This "world model" approach dramatically reduces the physics hallucinations that plague other video generation models.