2026-10-25 12:02

Introduction: Single RTX 4090 enables local inference, supporting dual control via text and camera. What is Hy-Nomas? Hy-Nomas is an open-source interactive video world generation model from Tencent Hunyuan.

What is Hy-Nomas?

Hy-Nomas is an open-source interactive video world generation model developed by Tencent Hunyuan, leveraging a segmented autoregressive architecture to generate videos. It supports seamless switching between text prompts and camera trajectory controls across segments, while maintaining long-term consistency through a lifelong KV cache. The code and model weights are fully open-sourced, offering distilled and INT8 quantized variants optimized for local inference on a single RTX 4090.

Key Features of Hy-Nomas

• Free-form Text Control: Supports arbitrary text prompts to govern objects, styles, scenes, and actions within the generated world.
• Camera Trajectory Control: Enables conditional generation based on camera motion paths, freely combinable with text prompts.
• Multi-turn Interactive Generation: The web-based interface allows injection of new instructions at segment boundaries for continuous, iterative generation.
• Prompt Enhancement (Nomas-PE): An integrated online model automatically converts informal user commands into structured generation prompts.
• Consumer-grade Local Deployment: Optimized with INT8 quantization and CPU offload, enabling inference on a single RTX 4090.
• LoRA Fine-tuning: Supports fine-tuning with custom video data, with resulting LoRA weights mergeable into the distilled model for inference.

Technical Principles of Hy-Nomas

• Segmented Autoregressive Architecture: Long videos are generated in segments, each conditioned on current prompts or camera signals, with a lifelong KV cache maintained across all segments to ensure temporal consistency in scenes, objects, and styles—enabling long-range world generation.
• Distillation for Accelerated Inference: Through distillation training, denoising steps are reduced to 4/6/8, paired with a specially fine-tuned PixelUMM decoder handling 32-channel DCAE-V latent variables, significantly reducing inference latency without compromising quality.
• Low VRAM Inference Mechanism: DiT weights and KV cache use INT8 storage with BF16 computation; historical context is truncated via sliding window to prevent unbounded growth; text encoder, DiT, and VAE are dynamically offloaded between CPU and GPU to keep peak VRAM usage within consumer-grade GPU limits.
• LoRA Segment-forcing Training: Fine-tuning starts from a non-distilled checkpoint using a segment-forcing training pipeline, training only LoRA parameters (rank/alpha = 256). After training, LoRA weights are merged back into the distilled checkpoint, yielding a model that combines distillation speed with customization capability.

Follow WeChat and reply “open-source” to join the AI Open-Source Project Discussion Group

Core Advantages of Hy-Nomas

• Dual-Modal Free Switching: Each segment can seamlessly switch between text prompt and camera trajectory control, enabling flexible interaction.
• Long-Term Consistency: Lifelong KV cache ensures object, scene, and style continuity across segments, preventing drift in long videos.
• Consumer-Grade Feasibility: INT8 quantization + sliding window + CPU offload enables local inference on a single RTX 4090.
• Ultra-Fast Inference: Post-distillation denoising requires only 4–8 steps, drastically reducing generation latency.
• Fully Open-Source: Code, weights, training, and data annotation pipelines are fully public under Apache 2.0 license for commercial use.

Project Links for Hy-Nomas

• Official Website: https://hy-nomas.hy.tencent.com/
• GitHub Repository: https://github.com/Tencent-Hunyuan/Hy-Nomas
• HuggingFace Model Hub: https://huggingface.co/tencent/Hy-Nomas

Competitive Comparison with Similar Models

Application Scenarios of Hy-Nomas

• AI Game Prototyping: Rapidly generate interactive game worlds via text and camera commands, enabling low-cost validation of gameplay concepts.
• Robotics Simulation Training: Leverage official Robotics LoRA to generate realistic robotic arm operation environments, providing synthetic training data for embodied intelligence.
• Film Previsualization & Storyboarding: Generate coherent long takes with cinematic LoRA to assist directors in shot composition and visual pre-review.
• Virtual Tourism & Spatial Exhibitions: Use camera trajectory control to generate navigable scenic spots, exhibition halls, and real estate environments, replacing traditional 3D modeling.
• AI Research & Education: Fully open-sourced training, fine-tuning, and data annotation code serves as a foundational platform for teaching and research in world modeling.

Source: AI Tools Kit