Repolode
Submit
Tencent-Hunyuan

Tencent-Hunyuan/Hy-Nomas

Hy-Nomas is a free-form interactive world generation model that builds video worlds from multi-turn text prompts and camera inputs.

View on GitHub
Stars
516
Forks
49
Open issues
1
Last commit
2 days ago

What is it?

Hy-Nomas is a video world generation model from Tencent Hunyuan. Users steer the generated world with multi-turn text prompts or camera trajectories, and each new segment continues from the previous ones using a persistent KV cache for consistency.

The repository provides training, inference, an interactive web demo, and data annotation code. Inference uses a distilled DiT with 4 to 8 denoising steps and a PixelUMM decoder for DCAE-V latents. A quantized INT8 checkpoint and FP8 decoder make single NVIDIA RTX 4090 inference possible with sliding-window history and CPU offload.

LoRA fine-tuning on custom video datasets is supported, and a captioning harness prepares training data from local videos. The project also has community integrations for SGLang, ComfyUI, and WanGP, and offers a Diffusers pipeline example.

Key features

  • 01Generates video worlds segment by segment from multi-turn text prompts.
  • 02Accepts camera trajectories as a control signal alongside text.
  • 03Runs single-GPU inference on an RTX 4090 with INT8 weights and FP8 decoding.
  • 04Includes a web demo for interactive exploration across turns.
  • 05Supports LoRA fine-tuning on custom video datasets.

Use cases

  • Prototyping interactive, prompt-driven video worlds or game-like environments.
  • Fine-tuning a LoRA adapter on your own video dataset for a specific visual style or domain.
  • Researching autoregressive video world models with camera and text control.
  • Generating robotics-style or cinematic scenes by merging the provided LoRA adapters.

Quick start

  1. 1

    Pull the pre-packaged Docker image

    docker pull hynomasoss/hynomasoss:cu130
  2. 2

    Start the container from the repository root

    docker run --rm -it --gpus all --ipc=host --network=host -v "$PWD:/workspace/Hy-Nomas" -w /workspace/Hy-Nomas hynomasoss/hynomasoss:cu130 bash
  3. 3

    Download the model weights from Hugging Face

    hf download tencent/Hy-Nomas --local-dir /path/to/Hy-Nomas
  4. 4

    Run the inference script from the README (set model paths and prompt file first)

Things to know

  • Requires a Linux host with a compatible NVIDIA GPU and CUDA 13.0; the README describes single RTX 4090 inference.
  • Model weights and attention kernel builds are large and can take considerable time to set up.
  • The metadata license is unclear (NOASSERTION), so check the LICENSE file before commercial use.

In collections

AI video generationAI video & audio tools

Topics

#video-generation#world-model#diffusion#lora#text-to-video#camera-control

FAQ

What is Hy-Nomas?
Hy-Nomas is a free-form interactive world generation model that builds video worlds from multi-turn text prompts and camera inputs. Hy-Nomas is a video world generation model from Tencent Hunyuan. Users steer the generated world with multi-turn text prompts or camera trajectories, and each new segment continues from the previous ones using a persistent KV cache for consistency.
Is Hy-Nomas free to use? What license does it have?
Hy-Nomas is open source under the Other license. Check the license terms before using it in commercial products.
Is Hy-Nomas still maintained?
Actively maintained. The last commit to Hy-Nomas was on Nov 6, 2026.
What can I use Hy-Nomas for?
Prototyping interactive, prompt-driven video worlds or game-like environments. Fine-tuning a LoRA adapter on your own video dataset for a specific visual style or domain. Researching autoregressive video world models with camera and text control. Generating robotics-style or cinematic scenes by merging the provided LoRA adapters.
How do I install Hy-Nomas?
Pull the pre-packaged Docker image: docker pull hynomasoss/hynomasoss:cu130 Start the container from the repository root: docker run --rm -it --gpus all --ipc=host --network=host -v "$PWD:/workspace/Hy-Nomas" -w /workspace/Hy-Nomas hynomasoss/hynomasoss:cu130 bash Download the model weights from Hugging Face: hf download tencent/Hy-Nomas --local-dir /path/to/Hy-Nomas Run the inference script from the README (set model paths and prompt file first)
Where can I get Hy-Nomas?
The source code and installation instructions are on GitHub: https://github.com/Tencent-Hunyuan/Hy-Nomas

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects