Eigen RadarAI
Archive
Analysis

Hy-Nomas lets users redirect a video world as it develops

Tencent’s Hy-Nomas accepts fresh text instructions and camera movements while generating successive video segments. Users can change the scene, objects and character actions, while a cache carries context from earlier segments. Code and weights are distributed under Apache 2.0. A separate single-card configuration also uses CPU memory, and an accompanying evaluation dataset provides scenarios for testing interactive control.

Artificial Intelligence··Midday
In an imagined stone courtyard, a cinema camera on rails faces a large tree on a wheeled platform; nearby paving slabs rest on separate carriages.

New instructions redirect successive video segments

Hy-Nomas, introduced by Tencent’s Hunyuan AI team, accepts text instructions and camera controls while generating successive video segments. Users can redirect the scene as generation progresses.[1], [2]

Instead of completing an entire video in one operation, the model continues segment by segment. Fresh instructions can add objects, change the visual style or environment and describe a character’s actions. Camera trajectories form another part of that control. Tencent’s October 23 announcement describes this interactive generation process.[1]

A cache carries context between segments

Information from earlier segments stays in a cache for subsequent generation. Tencent describes this as world memory intended to preserve context during long-running generation. Components include a Qwen text encoder, camera-conditioning elements and a PixelUMM video decoder. The company also lists lightweight LoRA adaptation for areas such as robotics, adding a limited set of weights to adapt a model for a specialized use.[1]

Control scenarios accompany the model

Code and weights are distributed under Apache 2.0. A separate single-NVIDIA-GeForce-RTX-4090 configuration uses INT8 and FP8 number formats alongside CPU memory. The Hy-Nomas-Bench dataset contains 280 test scenarios and 220 reference camera trajectories. Scenarios cover text instructions, camera movements and combinations of the two. The dataset contains neither videos nor model scores.[1]

References

  1. News sourceShiftDelete.NetTencent introduces Hy-Nomas for directing a video world as it develops↩1↩2↩3↩4
  2. News sourceAI時短ラボTencent Hy-Nomas generates video worlds under text and camera control↩
AI safety and regulationChip wars and the supply chain

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →