Description
Hy-Nomas is an advanced video world generation model designed to facilitate free-form, multi-turn interaction through text instructions and camera control. This model generates video segments one at a time, allowing users to introduce objects, change styles, and control scenes and camera angles as the generation progresses. The model architecture integrates a segment-autoregressive diffusion transformer (DiT), a Qwen text encoder, camera-conditioning modules, and a PixelUMM video decoder. A persistent key/value cache maintains context between segments, ensuring continuity in world generation.
Read More
The model supports flexible deployment options, including a standard sequence-parallel configuration and a single-GPU configuration, making it accessible for various hardware setups. Users can interact with the model through a web demo, which showcases its capabilities in real-time. Hy-Nomas also includes Hy-Nomas-Bench, a suite of benchmarks designed to evaluate interactive controllability and consistency during continuous world generation.
Hy-Nomas performs diffusion in a compressed video latent space and autoregression across video segments. Text, camera motion, and cached history condition the generation of each new segment, allowing for dynamic and interactive video creation. The model's architecture features 54 dual-stream MMDiT blocks, with text and video interacting through segment-wise causal joint attention.
The model's video latents are represented as DCAE-V latents, and the default inference pipeline decodes them with a fine-tuned PixelUMM decoder. This setup allows for efficient video generation with high-quality output. Hy-Nomas is licensed under Apache-2.0, and users can access the model and its components through the Hugging Face platform.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud, Self-hosted
- PlatformsWeb
Hy-Nomas Highlights
Free-form interaction through text and camera control
Segment-autoregressive diffusion transformer architecture
Persistent key/value cache for context retention
Flexible deployment with standard and single-GPU configurations
Hy-Nomas-Bench for evaluating interactive controllability
Compressed video latent space diffusion
54 dual-stream MMDiT blocks
Fine-tuned PixelUMM video decoder
Getting Started with Hy-Nomas
Access model: Visit Hugging Face
Authenticate: Set up account
Set up environment: Prepare hardware and software
Integrate via API: Use provided endpoints
Optimize: Adjust settings for desired output
Hy-Nomas's Use Cases
- Interactive Video Creation
- Game Development
- Virtual Reality
- Film Production
- Educational Content










