Tencent Hunyuan releases Hy-Nomas, an open-source interactive video world model

Tencent Hunyuan has open-sourced Hy-Nomas, a model that generates a video world segment by segment and that you steer as it goes: with free-form text prompts (add an object, change the style, scene or action) or with camera moves, including keyboard control in the demo. A world memory keeps long sequences consistent, and inference is distilled to 8 steps to run on consumer GPUs, according to Tencent. The code (training, inference, web demo, data annotation) and weights are released under Apache 2.0 on GitHub and Hugging Face, with LoRA fine-tuning for specialised domains such as robotics. An online demo is available; the technical report is announced.

See original publication↗Tencent Hunyuan (GitHub officiel) · github.com