Tencent wants to turn video generation into interactive worlds with Hy-Nomas

Tencent releases Hy-Nomas, an open video model capable of generating interactive worlds controlled by text and camera movements.

An AI-generated video typically follows a pre-defined sequence. Changing a character's movement, shifting the viewpoint, or returning to a set often requires running a new generation. With Hy-Nomas, Tencent explores another approach: building a video environment that evolves over the course of interactions, while retaining the information necessary for scene continuity. This open-source 8-billion-parameter model allows users to successively intervene on the action, the camera, and the appearance of a generated universe. It produces sequences up to 60 seconds long in 720p and can run, in an optimized configuration, on a single NVIDIA RTX 4090 graphics card. A video that evolves with prompts

Hy-Nomas belongs to the category of world models, systems that seek to represent the evolution of an environment rather than merely produce a sequence of images. Its operation relies on progressive generation: the model creates a first video segment, then uses subsequent prompts to continue the scene based on what already exists.

The user can ask a character to perform an action, introduce a new object, modify the appearance of a location, or change the visual style without systematically starting over from scratch. Prompts can be phrased in natural language, over several turns of interaction, rather than limited to a predefined list of commands.

The demonstrations published by Tencent notably show a panda climbing stone pillars, a snow leopard moving through a palace, and a train passing through a glacial tunnel. Other examples explore universes inspired by video games, animation, and realistic live-action footage. These scenes serve to illustrate the variety of supported prompts and styles.

The difference with a classic text-to-video generator lies in the ability to modify the direction of the sequence during its creation. Each new request fits into an already started environment, even as the model continues to synthesize images rather than running a traditional 3D engine. Directing actions and camera movements within the same environment

The system combines two forms of control. The first relies on textual descriptions: the user indicates what should appear or happen in the scene. The second uses camera trajectories to determine viewpoint changes and movements within the generated environment.

These commands can be alternated within the same sequence. It thus becomes possible to request an action, then move the camera to explore the set, and subsequently return to text prompts to evolve the characters or the environment.

On a technical level, Hy-Nomas represents trajectories using matrices that describe the camera's position and orientation. The model integrates them into its generation process to guide the viewpoint of subsequent frames. This method offers a more explicit spatial control than a simple prompt such as "move to the left" or "pan around the character."

However, the control remains generative. Hy-Nomas does not provide a fully reconstructed three-dimensional environment with editable objects and geometry directly usable in a game engine. It produces a video representation conditioned by the requested movements. A memory designed to preserve sets

Continuity is one of the major difficulties of interactive video models. When a camera leaves a location and then returns to it, previously visible elements may have changed shape, position, or appearance. To limit this drift, Tencent uses a generation memory based on a Key-Value cache, known as a KV cache.

This mechanism retains information from previous segments to make it accessible when creating subsequent ones. The model can thus rely on the video's history when it needs to represent a previously visited location. Tencent presents several demonstrations in which the camera executes a significant movement before returning to a set, a waterfall, or an architectural element.

Hy-Nomas features an advertised context window of 360,000 tokens, and its standard configuration retains the history of generated segments. However, this memory is not unlimited in practice: leveraging it increases hardware requirements as the sequence grows longer.

In the configuration designed for graphics cards with less memory, the system keeps a limited portion of the history on the GPU and archives older information in the system RAM. This compromise facilitates deployment, but it does not guarantee that every detail of a scene will remain perfectly identical after a long movement or numerous interactions. Eight billion parameters and segment-based generation

The architecture of Hy-Nomas combines a video diffusion transformer with a segment-organized autoregressive mechanism. The model uses 54 dual-stream multimodal blocks, responsible for processing text and video information, complemented by specialized modules for camera trajectories.

Prompts pass through an encoder based on Qwen2.5-VL-7B-Instruct. The images, for their part, are represented in a compressed latent space before being reconstructed by a tailored PixelUMM video decoder. This organization makes it possible to work on more compact representations than full images, while retaining the ability to produce a detailed sequence at the output.

Tencent also uses a distilled version of the model that reduces the generation process to eight denoising steps per segment in the documented configuration. The standard output is set to 720 × 1280 pixels, at 24 frames per second.

The project claims continuous generations reaching up to 60 seconds. However, this duration must be distinguished from computation speed: a video played back at 24 frames per second is not necessarily generated in real time. The published resources do not allow establishing a constant latency for each new prompt, regardless of the hardware used or the length of the history. A version tailored for consumer graphics cards

The Hy-Nomas GitHub repository includes a configuration specifically intended for inference on an NVIDIA GeForce RTX 4090. To reduce video