Hy-Nomas, Tencent's new open-source world model you can steer with text (huggingface.co)

Tencent has announced Hy-Nomas, an innovative open-source video world generation model designed for free-form, multi-turn interactions using text commands and camera controls. This groundbreaking technology enables users to generate and modify video environments in real-time—altering objects, styles, and scenes through descriptive text while controlling the viewpoint dynamically. The architecture of Hy-Nomas combines advanced techniques such as a segment-autoregressive diffusion transformer (DiT), a Qwen text encoder, and a dedicated PixelUMM video decoder, all orchestrated by a robust context-carrying KV cache that enhances continuity in world creation. This model holds significant implications for the AI/ML community, as it facilitates seamless and detailed user interaction, making it particularly appealing for applications in gaming, virtual reality, and content creation. Hy-Nomas is optimized for both standard and single-GPU configurations, supporting on-the-fly generation of high-resolution videos (720 × 1280 at 24 FPS) while leveraging advanced techniques like 8-step denoising and temporal compression in a compressed latent space. Furthermore, the introduction of Hy-Nomas-Bench—a suite of benchmarks measuring interactive controllability and consistency—extends the framework's utility by providing developers with a comprehensive set of tools for evaluating performance in real-time applications.