Tencent Hunyuan's interactive video world model, already indexed on 10 26: Videos are generated segment by segment, with each segment allowing text prompts to add objects, change styles, describe actions, or specify a camera trajectory, while previous content is continued via KV cache. This re-indexing was triggered by the subsequent release of the technical report and the addition of several community deployment options.
Rising score 69 · Likes 126 · Downloads 53401
Reasons
- New content compared to the 10 26 entry: The technical report was published on arXiv on 10 27 (2610.36729), stating that the backbone is a 8B diffusion model, using structured video descriptions combined with causal attention limited by action time windows, training via a multi-stage curriculum, and finally performing diffusion distillation. The abstract claims that human evaluation outperforms existing world models, but no specific scores are provided.
- Community integration: The third-party ComfyUI-Hy-Nomas node, which the author claims can run on an 12 GB GPU, allows starting a world from a single image or controlling it segment by segment using WASD keys; SGLang Diffusion has merged support, enabling both offline tasks and real-time streaming sessions; WanGP v18.81 has also added world generation and keyboard navigation.
- Sample provided by the ComfyUI node author: Four 5-second segments combined into 20 seconds, at 1280×704 resolution and 24 fps, taking approximately 3 minutes per clip on an RTX 4090. The official GitHub also states that the distilled version supports 4 / 6 / 8 step inference.
Usage
- ComfyUI: Search for Hy-Nomas (world model) in ComfyUI-Manager, or run
comfy node install hy-nomas. Downloadhy_nomas_distil_int8.safetensorsfrom PedroMarinhoDev/Hy-Nomas-ComfyUI and place it indiffusion_models, andhy_nomas_pixelumm_fp8.safetensorsinvae; use the text encoderqwen_2.5_vl_7b_fp8_scaled, which is the same as Qwen-Image. Set sampling to euler / simple / 8 steps / cfg 1, and choose a resolution of 1280×704 or 832×480. - Official code:
git clone Tencent-Hunyuan/Hy-Nomas. For single-GPU use, employconfigs/experiment/single_gpu_low_mem_inference.yaml, configured with INT8 DiT and FP8 PixelUMM; the standard configuration defaults to 8 GPUs for sequence parallelism. Prompts are in JSON format, specifying start and end times for each segment, with subsequent segments usingdescription_changeto define how the scene evolves. - If you do not want to deploy, you can first use the official online demo (hy-nomas.hy.tencent.com).
Risks
- The ComfyUI node is a community project, marked as "in development" by the author, so node inputs may change, and the code is licensed under GPL-3.0. The "start from image" approach works by placing the image into historical memory; the model itself does not have image input.
- The technical report abstract lacks quantitative scores, and the details of the human evaluation have not been verified by this column. The official single-GPU 4090 configuration requires substantial system memory to store offloaded models and accumulating historical cache, and the INT8 version cannot load LoRA during inference.
- The web demo requires a separate deployment of the prompt enhancement service Nomas-PE; 24 fps is merely the playback frame rate, and generation speed depends on hardware, resolution, and history length.