Hy-Nomas
An open-source world model from Tencent Hunyuan that uses a segmented autoregressive generation strategy and KV Cache to achieve continuous long-video generation. It comprises a language model, a DiT video diffusion transformer, and a Pixel UMM decoder, and supports text-prompt and camera-control modes.
Timeline (last 90 days)
Tencent's Hunyuan team open-sourced a world model named Hy-Nomas.
Hy-Nomas adopts a segmented autoregressive generation strategy, in which each later video segment is conditioned on the previous segment's content as context.
Hy-Nomas uses KV Cache as the memory mechanism for video context to maintain scene coherence.
The Hy-Nomas architecture consists of three main parts: a language model (Qwen2.5-VL-7B), a DiT video diffusion transformer, and a decoder (Pixel UMM).
The official demo includes a PE (prompt enhancement) step that converts natural-language prompts into structured descriptions, and this module can be deployed locally.
Hy-Nomas's official standard inference setup requires eight GPUs in parallel, and a single-GPU optimized version is also provided.
The uploader's hands-on testing found that Hy-Nomas's generation results are almost the same as the official demo, with the first 5-second video segment taking about 30 seconds to generate (based on high-performance GPU inference).
Hy-Nomas supports a camera control mode in which users can control the viewing direction with up, down, left, and right buttons, and the system records the operation trajectory and then generates video from the corresponding viewpoint.
When deploying Hy-Nomas locally, the uploader encountered problems such as flash-attention compilation failures and missing header files, and recommends using a ready-made image for deployment.
In the uploader's local test, the prompt enhancement module was not enabled and the resolution was about 720p, lower than the official demo, so generation quality may be slightly discounted.
1 more timeline events
All Facts (11)
Tencent's Hunyuan team already had prior work such as Hunyuan 3D and Hunyuan Video 1.1, and Hy-Nomas is its extension in the direction of spatial intelligence.
50%UnverifiedHy-Nomas adopts a segmented autoregressive generation strategy, in which each later video segment is conditioned on the previous segment's content as context.
50%UnverifiedHy-Nomas uses KV Cache as the memory mechanism for video context to maintain scene coherence.
50%UnverifiedThe Hy-Nomas architecture consists of three main parts: a language model (Qwen2.5-VL-7B), a DiT video diffusion transformer, and a decoder (Pixel UMM).
50%UnverifiedThe official demo includes a PE (prompt enhancement) step that converts natural-language prompts into structured descriptions, and this module can be deployed locally.
50%UnverifiedHy-Nomas's official standard inference setup requires eight GPUs in parallel, and a single-GPU optimized version is also provided.
50%UnverifiedThe uploader's hands-on testing found that Hy-Nomas's generation results are almost the same as the official demo, with the first 5-second video segment taking about 30 seconds to generate (based on high-performance GPU inference).
50%UnverifiedHy-Nomas supports a camera control mode in which users can control the viewing direction with up, down, left, and right buttons, and the system records the operation trajectory and then generates video from the corresponding viewpoint.
50%UnverifiedWhen deploying Hy-Nomas locally, the uploader encountered problems such as flash-attention compilation failures and missing header files, and recommends using a ready-made image for deployment.
50%UnverifiedIn the uploader's local test, the prompt enhancement module was not enabled and the resolution was about 720p, lower than the official demo, so generation quality may be slightly discounted.
50%UnverifiedTencent's Hunyuan team open-sourced a world model named Hy-Nomas.
50%