A video can change direction one segment at a time

Tencent has published Hy-Nomas, a model for generating video that can be steered across successive segments. Users can change the requested scene or action with text, or supply camera movements to guide the next part.

The public release is more than a gallery. Tencent’s code repository provides training, inference, demo and data-annotation code, while its model repository lists generation and decoding checkpoints.

This is interactive video generation, not evidence of a complete physical simulator. The immediate claim is that instructions and camera controls can shape a continuing sequence. How reliably the sequence preserves objects and spatial relationships is a question to test.

Remembering earlier segments has a resource cost

The model carries information from earlier segments in a cache, a store of intermediate values reused during generation. That supplies context for the next segment; it does not guarantee perfect memory or consistency.

The documentation distinguishes a standard configuration from a lower-memory, single-GPU configuration. The latter uses reduced-precision components, limits the active history on the GPU and moves parts of the workload through system memory.

Those transfers add latency, and the CPU-side history archive can keep growing. The single-GPU description is therefore a configuration claim with conditions, not a promise that any consumer computer can run the whole system smoothly.

Playback rate is not generation speed

The model card lists output at 24 frames per second. It explicitly says that this is the video’s playback rate. Generation speed depends on hardware, resolution, segment length, history and configuration.

The local web demo also uses Nomas-PE, a separate service that turns user instructions into generation prompts. The documentation says that service needs its own serving resources; the single-GPU claim applies to video inference.

This matters when comparing demos. A clip that plays fluidly does not reveal how long it took to generate, and a model’s inference requirement is not necessarily the resource requirement of the full application.

Published tests make evaluation possible, not automatic

Tencent also provides Hy-Nomas-Bench: 280 cases divided among multi-turn text control, camera navigation, interleaved text and camera input, and spatial memory. The spatial-memory task tests revisits to a distinctive target, which makes consistency a specific question rather than a visual impression.

The dataset includes prompts, interaction sequences and camera trajectories. Its card says generated videos, extracted reference or revisit frames, and evaluation scores are not included.

Model Current has not run the model or reproduced its evaluations. The release gives researchers materials to examine control and continuity, but the availability of those materials should remain separate from claims about measured quality or real-time operation.

Sources

  1. Tencent: Hy-Nomas model card and assetsPrimary model repository: segment generation, cache and low-memory caveats, playback-versus-generation distinction and separate Nomas-PE serving needs. Public file listing inspected; large weights not downloaded or executed.
  2. Tencent-Hunyuan: Hy-Nomas codePrimary public code repository provides training, inference, demo and annotation material. Technical report is marked coming soon; no peer-review claim or independent replication inferred.
  3. Tencent: Hy-Nomas-BenchPrimary dataset card: 280 cases in four test splits, interaction/control tasks and trajectories. Generated videos, reference/revisit frames and evaluation scores are explicitly absent.
  4. Tencent-Hunyuan: full web-demo deploymentPrimary application guide explicitly requires Nomas-PE and its separate GPU resources even in single-GPU video mode. Application readiness is distinguished from a gateway health response. No environment or model launched.