HY-World 2.0 just dropped
Posted by bobeeeeeeeee8964@reddit | LocalLLaMA | View on Reddit | 15 comments
Tencent just released HY-World 2.0 — the first open-source state-of-the-art 3D world model.
What makes it special (repo’s own words):
Generates real 3D assets (3D Gaussian Splats, meshes, point clouds) — not videos
Persistent & editable worlds you can “build once, keep forever”
Fully importable into Unity, Unreal Engine, Blender, Isaac Sim
Native physics, collision, first-person navigation, real-time rendering on consumer GPUs
Unlike video-only models (Genie 3, HY-World 1.5), no flickering, no time limit, true 3D consistency
Core: WorldMirror 2.0 — unified feed-forward model that predicts depth, normals, camera params, point clouds + 3DGS in one pass.
Multimodal: text, image, multi-view, or video → 3D world.
HF: https://huggingface.co/tencent/HY-World-2.0
GitHub: https://github.com/Tencent-Hunyuan/HY-World-2.0
DHasselhoff77@reddit
If you look at the video full screen you can see that both the texture and mesh resolution are very low. Still interesting results.
Long_comment_san@reddit
"no flickering, no time limit, true 3D"
Right...
Kodix@reddit
This seems *really cool* at a glance. Especially since it's FOSS, apparently? To the best of my knowledge, all of the current 3D asset generation models that actually work well have been closely guarded.
Only issue I see is that, at the moment, this generates entire *scenes*, rather than individual models that make up those scenes. In that cozy fireplace scene it generated, you can't open that up with Unity and, say, move a table easily.
Still, cautiously exciting.
no-adz@reddit
"We will release all model weights, code, and technical details to facilitate reproducibility and advance research in this field."
Open source, not free license
FriskyFennecFox@reddit
No, unfortunately not FOSS
Kodix@reddit
You're right, I'm *horribly* wrong. I just glanced at the claims they make in the model cards instead of reading the license. It's very much limited.
gK_aMb@reddit
My 3D model benchmark.
akumaburn@reddit
Another great open source contribution by a Chinese company.. meanwhile in the west..
redditscraperbot2@reddit
Some BIG asterisks here. The code available is for making Gaussian splats from images and videos.
Anyone downloading this and thinking they’ll get gta6 no mistakes is going to be really disappointed
Finanzamt_Endgegner@reddit
It also allows text as inputs.
redditscraperbot2@reddit
Unless I’m mistaken the current world mirror script only takes images, videos and some 3D formats. Text is for the panorama scripts coming sooner or late.
Finanzamt_Endgegner@reddit
Hmm in that case the description in the repo is kinda misleading which might be why you got down voted.
redditscraperbot2@reddit
Yeah I spent all morning tinkering with it. Like the splats are really good for the information you give them. But they’re just splats.
Finanzamt_Endgegner@reddit
Good to know 👍
FinBenton@reddit
Wish they had a demo hosted, apparently this release is just part of it and not the complete system.