What happened
Tencent released Hy-Nomas, a video world generation model, on Hugging Face in a BF16 Diffusers format. The model enables free-form, multi-turn interaction where users can introduce objects, change styles, or control the camera as video segments are generated sequentially. The release includes Hy-Nomas-Bench, a suite of four benchmarks with 280 test cases designed to evaluate interactive controllability and spatial memory in continuous world generation.
Tencent published the Hy-Nomas model on Hugging Face, specifically providing a distilled version in BF16 Diffusers format. The model is designed for video world generation, allowing users to interact with a generated environment through text instructions and camera control. Unlike traditional text-to-video models that generate a single fixed clip, Hy-Nomas generates the world one video segment at a time, enabling continuous interaction.
The architecture combines a segment-autoregressive diffusion transformer (DiT), a Qwen text encoder, camera-conditioning modules, and a PixelUMM video decoder. A persistent key/value (KV) cache carries context between segments, ensuring consistency as the user navigates or modifies the scene. The text encoder uses the language-model component of Qwen/Qwen2.5-VL-7B-Instruct, while the video decoder is a fine-tuned PixelUMM transformer.
The release also includes Hy-Nomas-Bench, a benchmark suite comprising four benchmarks with 280 test cases. These benchmarks cover multi-turn free-form text interaction, camera navigation, interleaved text and camera control, and spatial memory. The suite provides structured prompts, interaction sequences, and reference camera trajectories to evaluate how well the model maintains consistency and controllability during continuous generation.
The model is licensed under Apache-2.0, allowing for open-source use and modification. The Hugging Face repository contains the BF16 Diffusers conversion, while the original model repository provides native checkpoints, quantized variants, and adapters. The release notes indicate that the Diffusers pipeline accepts plain text or prepared structured prompts directly, without requiring the separate Nomas-PE prompt enhancement service used in the original web demo.
Source details: huggingface.co ↗
Why it matters
This release provides a concrete, open-source implementation of interactive video world generation, moving beyond static text-to-video synthesis. By supporting persistent context through a key/value cache and explicit camera conditioning, it allows for dynamic, user-directed exploration of generated environments. The inclusion of a dedicated benchmark suite addresses the lack of standardized evaluation metrics for interactive consistency and spatial memory in generative video models.
Hy-Nomas represents a significant step in the development of interactive generative media by enabling real-time, multi-turn control over video generation. The ability to introduce objects, change styles, and control the camera dynamically addresses a key limitation of current video models, which often lack the ability to respond to user input after the initial prompt is processed.
The technical approach of using a persistent KV cache for context and explicit camera conditioning modules provides a robust framework for maintaining spatial and temporal consistency in generated worlds. This is crucial for applications in gaming, virtual reality, and interactive storytelling, where user agency and environmental coherence are paramount.
The introduction of Hy-Nomas-Bench is particularly noteworthy as it provides a structured way to evaluate the specific challenges of interactive video generation, such as spatial memory and camera navigation. This fills a gap in the current evaluation landscape, which has largely focused on static video quality metrics rather than interactive controllability.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Developers should monitor the integration of Hy-Nomas into broader creative tooling and the adoption of Hy-Nomas-Bench as a standard for evaluating interactive video models. The availability of the model under the Apache-2.0 license suggests potential for rapid community adaptation and .
The adoption of Hy-Nomas-Bench by other research groups and industry players will determine its impact as a standard for evaluating interactive video models. If widely adopted, it could drive further innovation in this specific area of .
Community contributions and of the Hy-Nomas model are likely to emerge given its Apache-2.0 license and availability in a standard Diffusers format. This could lead to specialized versions of the model for specific use cases, such as game asset generation or virtual environment design.
The performance of Hy-Nomas in real-world applications, particularly in terms of latency and resource requirements, will be a key factor in its practical adoption. The release notes mention the use of specific attention kernels for optimal performance, which may influence how developers deploy the model.