August 12 News.Kunlun Wanwei SkyWork AI Technology Release Week was launched on August 11, with one model released every day for five consecutive days, covering core multimodal AI scenarios.
Today, Kunlun World Wide brings the upgraded version of the Matrix-Game interactive world model in the Matrix series of self-developed world models --- the Matrix-Game interactive world model."Matrix-Game 2.0"., claiming to be a world model that realizes interactive real-time long sequence generation in generic scenarios.
and in order to promote the field of interactive world modeling."Matrix-Game 2.0" Full Open Source, which is claimed to be the industry's first open source solution for real-time long sequence interactive generation of world models on generic scenarios.
"Matrix-Game 2.0".It is claimed to be a "quantum leap" in real-time generation and long sequence capabilities. Compared to the previous version, version 2.0 focuses more on low-latency, high-frame-rate long sequence interactive performance.Stable generation of continuous video content at 25 FPS in a wide range of complex scenarios, with scalable generation durations down to the minute level, dramatically improving coherence and utility.
While the reasoning speed is significantly increased, the model still maintains a "precise understanding" of the physical laws and scene semantics, and supports users to freely explore, manipulate, and build real-time virtual environments with clear structure, rich details, and reasonable rules through simple commands.
Kunlun Wanwei also open-sources Matrix-3D Large ModelThe result is a high-quality, trajectory-consistent panoramic video that starts from a single image and directly restores the roamable 3D space, benchmarking against Feifei Li's WorldLabs generation for a larger range of explorable spaces.
Matrix-3D consists of the following core components:
- Trajectory-guided panoramic video generation module:Using scene mesh renderings as conditional inputs, a video diffusion model is trained to generate panoramic videos that conform to a given camera trajectory. It improves the consistency of the generated video in terms of spatial structure and mitigates occlusion errors and image artifacts.
- Dual-path selectable panoramic 3D reconstruction module:Optimization Path: Optimize the generated video with super-segmentation and 3DGS to obtain high-quality 3D structures. Feedforward network path: based on Transformer direct regression, fast prediction of 3D geometric attributes from Latent features of the generated video, realizing efficient reconstruction.
- Matrix-Pano dataset:Large-scale, high-quality synthetic dataset containing 116K static panoramic video sequences with camera tracks, depth maps, and text annotations.
1AI attaches the open source addresses of the two models as follows:
https://github.com/SkyworkAI/Matrix-Game
https://github.com/SkyworkAI/Matrix-3D