Yes, the code and model weights are released as an open-source research project, available for free on GitHub.

How long can LongVie 2 videos be?

The model is designed to generate consistent video sequences of up to 3 to 5 minutes, which is significantly longer than the 5 10 second clips typical of models like Sora or Runway Gen 2.

Can I control what happens in the video?

Yes, LongVie 2 is "controllable," meaning you can feed it depth maps or skeleton data (pose estimation) to dictate exactly how a character moves or how a scene is laid out, rather than just relying on a text prompt.

What hardware do I need?

As a research grade "World Model," it requires significant VRAM. While specific consumer requirements vary, running inference on minute long videos generally requires enterprise grade GPUs (like NVIDIA A100 or H100) or multi GPU setups.

LongVie 2

Multimodal Controllable Ultra-Long Video World Model.

From Shanghai AI Laboratory, Fudan University, NVIDIA, NTU (Research Collaboration)

Founder: Jianxiong Gao, Zhaoxi Chen, Xian Liu, et al. (Research Team)

Video & Animation

Dec 2025

10K+

Free

Pricing Model

Free

Starting Price

$0/Month

About This AI

LongVie 2 is an advanced “Video World Model” designed to solve the biggest problem in AI video duration.

While most models struggle to generate coherent clips longer than 10 seconds, LongVie 2 uses a novel three stage training process to generate ultra long videos lasting 3 to 5 minutes with high consistency.

It integrates “dense” controls (like depth maps) and “sparse” controls (like keypoints) to give creators precise direction over the video’s content, ensuring that characters and environments don’t “hallucinate” or morph into different objects over time.

AI Art Image To Video Text To Video Video Generator

Pricing

Pricing Model

Free

Starting Price

$0/Month

Key Features

Ultra Long Generation Capable of generating coherent video sequences lasting 3, 5 minutes (approx. 3000-5000 frames) without losing quality.
Multimodal Control Accepts both dense inputs (depth maps) and sparse inputs (skeleton keypoints) to guide character motion and scene structure.
Degradation Aware Training Uses a specialized training method that anticipates and corrects the "blur" and artifacts that usually accumulate in long AI videos.
History Context Guidance Maintains a memory of past frames to ensure that an object generated in minute 1 still looks the same in minute 3.
LongVGenBench Introduced a new benchmark specifically for testing the limits of long form video generation.
Autoregressive Architecture Builds the video frame by frame (or chunk by chunk) allowing for infinite potential extension.

Pros

Breaks the "10 second barrier" of standard AI video tools.
High level of control over character movement via keypoints.
Completely free and open source for researchers.
Reduces "drifting" (where the scene changes randomly) significantly.
Backed by top research institutions (NVIDIA, Shanghai AI Lab).

Cons

Requires substantial GPU power (likely H100s) for efficient inference.
Setup is technical (requires Python, PyTorch, Linux).
Current resolution in research demos is often limited (e.g., 352×640) to save compute.
No user friendly "app" yet; it is a codebase for developers.

Best for AI Researchers, Computer Vision Engineers, and Technical Artists looking to experiment with long-form storytelling and coherent video synthesis.

FAQs

Is LongVie 2 free?
Yes, the code and model weights are released as an open-source research project, available for free on GitHub.
How long can LongVie 2 videos be?
The model is designed to generate consistent video sequences of up to 3 to 5 minutes, which is significantly longer than the 5 10 second clips typical of models like Sora or Runway Gen 2.
Can I control what happens in the video?
Yes, LongVie 2 is “controllable,” meaning you can feed it depth maps or skeleton data (pose estimation) to dictate exactly how a character moves or how a scene is laid out, rather than just relying on a text prompt.
What hardware do I need?
As a research grade “World Model,” it requires significant VRAM. While specific consumer requirements vary, running inference on minute long videos generally requires enterprise grade GPUs (like NVIDIA A100 or H100) or multi GPU setups.