Tag: #3d-generation40 results
Researchers from PKU, CUHK, and Shanghai AI Lab introduced VGGT-Edit, a framework that enables 3D scene editing in just 5 seconds. This represents a massive 120x performance improvement over previous methods.
ByteDance launched Seed3D 2.0, a next-generation 3D foundation model. It sets new benchmarks in 3D generation capabilities. The model leads in both geometry and texture quality.
Alibaba Group launched Happy Oyster, an AI model that generates interactive 3D video environments. This expands AI capabilities from 2D video generation into immersive content creation.
Tencent released and open-sourced HY-World 2.0, its 3D world model. Users can generate interactive 3D environments from text or images. It integrates with major game engines.
AI leaders like Yann LeCun's AMI and Fei-Fei Li's World Labs secure over $2B in funding for world models to overcome LLM limitations in physical understanding. Key approaches include JEPA for abstract predictions (V-JEPA 2), spatial reconstruction (Marble), and simulation (Genie 3). These routes aim to enable AI with true 3D, causal, and interactive world comprehension.
Zhiyuan's Genie Sim 3.0 simulation platform receives major upgrade for full-cycle embodied AI development. It enables instant high-fidelity 3D scene generation from text or single images. Environment build time drops from hours to minutes.
VAST CEO Cao Yanpei discusses how 2-second 3D generation represents the true speed standard. This signals the arrival of the AI 3D 2.0 paradigm, promising faster workflows.
Tencent Hunyuan 3D team has open-sourced WorldCompass, the first reinforcement learning post-training framework for world models. It acts as a 'compass' for long-sequence interactive models, guiding accurate user instruction following and visual consistency. Designed specifically for extended temporal and interactive world model training.
AI startup VAST secured $50 million in Series A funding. It unveiled a new generation of 3D foundation models continuing to set industry SOTA. The company is preparing to launch its first world model.
MIT's PhysiOpt optimizes 3D generative models directly in latent space using differentiable physics, treating implicit fields as density distributions. Avoids costly remeshing, enabling editable, structurally sound designs. Accepted to SIGGRAPH Asia 2025.