Metavershche and AIshcheTencent presented the open AI model Hy-Nomas for interactive videos
Tencent has unveiled Hy-Nomas, an open-source artificial intelligence model for creating interactive video worlds with context retention. The architecture enables the generation of sequential scenes, camera control, and plot modification via text commands.
Memory retention between video segments is ensured through KV-cache technology. This allows generation to continue by taking into account previous events in the history.
The model is based on a Segment-autoregressive MMDiT with 54 blocks. Text processing uses the Qwen2.5-VL-7B encoder, while video compression is implemented via DCAE-V and PixelUMM.
The output video has a resolution of 720p and a frame rate of 24 frames per second. Generation takes 8 diffusion steps. Cinematic and Robotics LoRA adapters are available.
Developers have optimized the model launch for consumer graphics cards. It can run on a single RTX 4090 with 24 GB VRAM thanks to INT8/FP8 quantization and CPU offload.
Limiting active history and parameter quantization help reduce resource requirements. However, generation is not fully realistic in terms of time.
The project is distributed under the Apache 2.0 license. Integrations with ComfyUI, SGLang, WanGP, and Reactor have already appeared. Third-party developers are creating applications for interactive mini-dramas.
