Tencent Hunyuan Hy-Nomas: Interactive Video World Model Architecture and Setup
Hy-Nomas is an open-source interactive video world model controlled by text prompts and camera trajectories, with single RTX 4090 low-memory inference and LoRA
Hy-Nomas is Tencent Hunyuan's open-source interactive video world generation framework controlled by multi-turn text prompts and camera trajectories. Its GitHub repository was created on October 21, 2026, and Tencent Hunyuan introduced the project publicly on October 24. The repository provides tools for segment-based generation, lifelong KV caching for visual consistency, and single-GPU execution.
Image source: https://github.com/Tencent-Hunyuan/Hy-Nomas
Hy-Nomas is an interactive video world generation model supporting free-form multi-turn text prompt control or camera control. The repository provides training, inference, interactive demo, and data annotation code.
Segment-Based Autoregressive Architecture and Lifelong Memory
The technical foundation of Hy-Nomas centers on an autoregressive, segment-wise generation pipeline.
- Persistent World Memory: The model accepts segment-level text prompts or camera signals while maintaining a lifelong KV cache for consistency.
- Free-Form Text Prompting: Free-form natural language inputs can generate objects, styles, scenes, and actions.
- Trajectory Conditioning: Camera trajectories condition generation alongside text prompts.
Single RTX 4090 Low-Memory Inference and Distillation
Hy-Nomas integrates architectural optimizations aimed at consumer workstation GPUs.
- Distilled Denoising: The inference pipeline uses 4 to 8 denoising steps.
- DCAE-V Latent Decoder: The model uses 32-channel DCAE-V latents and a fine-tuned PixelUMM decoder.
- Consumer Hardware Target: A low-memory inference path supports a single RTX 4090.
Installation Workflow and LoRA Fine-Tuning Pipeline
The repository provides end-to-end tooling for deploying interactive demos and adapting weights to downstream domains.
Setup is recommended via a pre-packaged Docker image (hynomasoss/hynomasoss:cu130) containing the Python environment, PyTorch, and required CUDA attention kernels. Alternatively, a manual installation path is provided for Linux environments using Python 3.12+ and uv.
For specialized domains such as robotics, Hy-Nomas includes video annotation tools, offline sequence packing scripts, and a LoRA (Low-Rank Adaptation) fine-tuning pipeline, along with code for running a browser-based interactive demo.
Sources
- GitHub: Tencent-Hunyuan/Hy-Nomas