Hy-Nomas
Hy-Nomas generates a world through consecutive video segments. Text instructions and camera controls can change objects, actions, scenes and visual style while generation continues.
Model design and access
A segment-wise diffusion transformer uses a persistent attention cache and a PixelUMM video decoder. A separate RTX 4090 configuration uses quantization and CPU offloading. It bounds active GPU history while retaining a CPU archive; total memory can still grow.Hy-Nomas official model documentation
Practical use
Practical guidance. Treat the result as generated video rather than a guaranteed geometric simulation. Check whether objects and room layouts remain consistent when returning to a previous viewpoint. Output frame rate is not the same as generation speed.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| License | Apache 2.0; third-party components retain their licensesHy-Nomas official model documentation |
| Reference output | 720 × 1280 at 24 frames per secondHy-Nomas official model documentation |
| Distilled sampling | 8 denoising stepsHy-Nomas official model documentation |
Related models
- HY-World 2.0
Tencent · Hy world models
A model for generating and reconstructing 3D scenes from multiple images or video.
Launch date not confirmedDownloadable - HY-WorldPlay
Tencent · Hy world models
A streaming world model that predicts action-conditioned video frames from text or a starting image.
Launch date not confirmedDownloadable