Hy-Nomas Operational

Hy-Nomas is an open-source interactive video world generation model developed by Tencent Hunyuan. It generates videos autoregressively in segments, supports freely switching between text prompts and camera trajectory control between segments, and uses a lifelong KV cache to maintain long-term consistency. Hy-Nomas has released its code and weights, including distilled and INT8-quantized versions. After optimization, it can run inference locally on a single RTX 4090.

Free-form text control: Supports text prompts in any format to control objects, styles, scenes, and actions in the generated world.
Camera trajectory control: Generates content conditioned on camera motion trajectories, which can be freely combined and switched with text prompts.
Multi-turn interactive generation: The Web interface supports injecting new instructions at segment boundaries for continuous multi-turn generation.
Prompt enhancement (Nomas-PE): An integrated online model automatically converts users' natural-language instructions into structured generation prompts.
Consumer-grade local deployment: INT8 quantization and CPU offload optimization enable inference on a single RTX 4090.
LoRA fine-tuning: Supports fine-tuning with custom video data and merging the result into the distilled model for inference.

Follow WeChat and reply with “开源” to join the AI open-source project group
Segmented autoregressive architecture: Long videos are generated segment by segment. Each segment is conditioned on the current prompt or camera signal, while a lifelong KV cache spanning all segments maintains temporal consistency in scenes, objects, and styles for long-range world generation.
Distillation-accelerated inference: Distillation training compresses denoising to 4/6/8 steps. Combined with a specially fine-tuned PixelUMM decoder for 32-channel DCAE-V latent variables, this substantially reduces inference latency while maintaining quality.
Low-memory inference: DiT weights and the KV cache are stored in INT8 while computation remains in BF16. Historical context is truncated with a sliding window to prevent unlimited growth. The text encoder, DiT, and VAE switch between CPU and GPU by stage, keeping peak VRAM usage within the range supported by consumer graphics cards.
LoRA segment-forcing training: Fine-tuning uses a non-distilled checkpoint as the base model and a segment-forcing training pipeline. Only LoRA parameters are trained (rank/alpha 256). After training, the LoRA weights are merged back into the distilled checkpoint to produce an inference model that retains distillation acceleration while supporting customization.

Flexible switching between two modalities: Each segment can freely switch between text prompts and camera trajectory control, enabling flexible interaction.
Long-term consistency: The lifelong KV cache preserves objects, scenes, and styles across segments, preventing drift in long videos.
Runs on consumer hardware: INT8 quantization, sliding-window context, and CPU offload enable local inference on a single RTX 4090.
Very fast inference: Distillation requires only 4 to 8 denoising steps, significantly reducing generation latency.
Fully open source: The code, weights, training process, and data annotation workflow are publicly available under the Apache 2.0 license for commercial use.

Official website: https://hy-nomas.hy.tencent.com/
GitHub repository: https://github.com/Tencent-Hunyuan/Hy-Nomas
Hugging Face model repository: https://huggingface.co/tencent/Hy-Nomas

AI game prototyping: Quickly generate interactive game worlds through text and camera instructions to validate gameplay concepts at low cost.
Robot simulation training: Use the official Robotics LoRA to generate realistic robotic arm operation environments and provide synthetic training data for embodied intelligence.
Film previsualization and storyboarding: Use cinematic LoRAs to generate coherent long takes, helping directors design camera movements and preview visuals.
Virtual cultural tourism and spatial showcases: Use camera trajectory control to generate explorable scenic areas, exhibition halls, and real estate scenes, replacing traditional 3D modeling.
AI research and education: The complete training, fine-tuning, and data annotation code is open source, providing an educational and research foundation for world model development.