Скачать или смотреть LongVie 2: Controllable Ultra-Long Video Gen

In this AI Research Roundup episode, Alex discusses the paper: 'LongVie 2: Multimodal Controllable Ultra-Long Video World Model' This work turns short-clip diffusion video backbones into a unified "video world model" that can generate minute-scale videos with strong controllability and temporal consistency. The authors introduce LongVie 2, an end-to-end autoregressive framework with three progressive training stages. Stage I injects both dense (depth maps) and sparse (tracked 3D keypoints/points) controls into a frozen diffusion backbone using a ControlNet-style design, carefully regularizing them so neither dominates. Stage II uses degradation-aware training, corrupting first frames to better match real inference, while Stage III enforces cross-clip coherence to preserve long-term visual quality across extended video horizons. Paper URL: https://arxiv.org/abs/2512.13604 #AI #MachineLearning #DeepLearning #VideoGeneration #DiffusionModels #Autoregressive #Multimodal

Resources:
GitHub: https://github.com/Vchitect/LongVie