用语言改剧情，用镜头探索：腾讯混元 Hy-Nomas 开源
Rewrite the Plot with Language, Explore with Camera: Tencent Hunyuan Hy-Nomas Open Source
Summary
Click “Computer Vision Life”, select the “Starred” robot AI content to receive updates instantly, make characters act, rotate the camera to observe, then use new commands to advance the story. Next, you can finish this round before deciding. Tencent Hunyuan open‑source H
Click “Computer Vision Life”, then select “Star”
Robot AI insights delivered instantly
Let the character act, rotate the camera to observe, and then use a new command to advance the story.Next step, you can finish this round before deciding.
Tencent Hunyuan open‑source Hy‑Nomas, which explores exactly this Free‑form interaction: specify actions with natural language, adjust the viewpoint with continuous camera control, and let subsequent generation incorporate new inputs while continuing the existing scene and story. The code, model weights, and evaluation benchmarks are open, and readers can also experience it themselves through the online demo.
Text drives the plot, the camera takes you exploring. In the same world, each turn decides what happens next.
Watch the promo video
01 Group Four Demo, Enter Free Interaction
One sentence decides what happens next
Let characters act, change the environment, and drive the story; free‑text control connects these ideas to video generation.Commands are not limited to predefined ones, you can issue the next request based on the current scene and see how the world responds.
|
|
|---|
Let words and lenses take turns, keep exploring
First issue a textual command, then move the camera to look around, and follow up with a new instruction. Hy-Nomas supportsalternating text commands with continuous camera control, allowing actions and observations to flow together in the same world.
Continue the story: generate a presentation in 60 seconds
The following three examples all demonstrate 60‑second video generation. As continuous commands progress, the scene and actions unfold step by step, allowing you to see how the preceding and following segments connect and how the story evolves with new inputs.
Leave and return, observing the world’s memory
Rotate the camera, leave a scene, and then return to the previous spot—does the object’s appearance and the environmental layout still match?continuous exploration of scene consistency. World memory is also part of the evaluation below, and there is still room for improvement.
The project also showcases different styles such as multi-angle, realistic, and anime, as well as examples of scene adaptation; more demos can be viewed on the project homepage.
02 280 cases, view interaction performance
Beyond the demo, the project validates interaction performance through 280 cases of Hy-Nomas-Bench to assess interaction performance, covering multi‑turn text interaction (60), camera control (100), alternating text and camera control (60), and spatial memory (60). Each case is executed consecutively in the same world without resetting the scene between rounds, and pairwise evaluations are performed by three annotators who have undergone standard calibration.
The figure below uses Hy-Nomas as a reference,Green, gray, and pink represent win, draw, and loss respectively. The GSB ratio reflects relative preferences in the evaluation and does not represent task success rate.
Each model uses its supported interaction interface; among them, LingBot-World v2 uses the adapted open-source model.
Key results to note in the evaluation:
-
Text command compliance: Compared to HappyOyster-Directing and the adapted LingBot-World v2, the win rates are **65.0% and 75.0%**, respectively. -
Text and camera alternating control: Compared to the adapted LingBot-World v2, the success rates for command compliance and physical plausibility are **75.0% and 73.3%**. -
Camera Control: All three comparisons are mainly draws, with win rates higher than loss rates. -
Spatial Memory: Compared to LingBot-World v2, the win rate is 58.3%, close to HappyOyster; compared to Genie 3, the win/draw/loss rates are 13.3%, 56.7%, and 30.0%.
These results demonstrate Hy-Nomas’s potential in text instruction response and alternating control, and also indicate the current limits:Long-range spatial consistency still has room for improvement.
03 Try it now
If you want to see how this world responds to your commands, you can first try the online demo; you can also download the model, view the code and benchmarks, and further explore your own interaction scenarios.
Experience and Demo
-
Online demo 👉 https://hy-nomas.hy.tencent.com/demo -
Project homepage 👉 https://hy-nomas.hy.tencent.com/ -
Hunyuan official site introduction 👉 https://hy.tencent.com/model/open-source-model/detail?id=hy-nomas
Open Source and Resources
-
Code and deployment instructions 👉 https://github.com/Tencent-Hunyuan/Hy-Nomas -
Model weights 👉 https://huggingface.co/tencent/Hy-Nomas -
Technical report 👉 https://github.com/Tencent-Hunyuan/Hy-Nomas/blob/main/Hy_Nomas_Tech_Report.pdf -
Evaluation benchmark 👉 https://huggingface.co/datasets/tencent/Hy-Nomas-Bench
The next command, the next perspective, it’s up to you to decide. If you’re also interested in generative worlds with free interaction, feel free to try Hy-Nomas and leave a Star, and follow the project’s progress.
New Arrivals
Recommended Reading
Detailed introduction! The ultra‑long‑range EDU‑SCAN 3D LiDAR scanner, born for an open ecosystem!
Major Release! Unlock Hundred-Meter-High Drone Mapping!
Major Release! Rebuild a Hundred-Meter-High Building in 2 Minutes!
Major release! Empowering robot dog SLAM and reconstructing the 3D world!
SLAM Challenge! EDU-SCAN clearly reproduces a huge indoor electronic screen
SLAM Challenge! Reconstruct 10,000 in 9 minutes
This research marvel smoothly reproduces the strongest SLAM: Fast-LIVO2!
Get MirrorSpeed
Get MirrorSpeed