HunyuanWorld 1.0

What it is

HunyuanWorld 1.0 is a 3D world generation model from Tencent's Hunyuan team, released on 26 July 2025 together with a technical report on arXiv ^[1]. It takes a text prompt or an image as input and produces an immersive 3D world rather than a single object or a flat picture. Tencent described it as the first open-source, simulation-capable, immersive 3D world generation model ^[1].

The output is not a locked-down video or a rendered image: scenes come with mesh export, and objects inside them are represented separately from one another, which is what makes interaction with individual objects possible ^[1].

How it works

A central design choice is the use of 360° panoramic proxies ^[1]. Instead of predicting 3D geometry directly, the model first produces a panorama that covers the full field of view around the viewer, and then builds the 3D world on top of that intermediate representation. Panoramas are a convenient middle step because they already wrap all the way around a scene, so the geometry can be lifted from a complete surround view rather than from a single narrow camera angle.

The technical report describes the framework as built around a semantically layered 3D mesh representation at its core ^[2]. In other words, the scene is held as a mesh whose parts carry semantic meaning, so that the ground, the objects and the surroundings are not one undifferentiated surface. Combined with the disentangled object representations, this is what allows a generated world to be edited and interacted with rather than merely viewed ^[1].

The open-source version is based on Flux ^[1].

Released checkpoints

The release included four checkpoints: PanoDiT-Text, PanoDiT-Image, PanoInpaint-Scene and PanoInpaint-Sky ^[1]. The names reflect their roles: two of them are conditioned on text and on images respectively, and two are inpainting modules, one aimed at the scene and one at the sky.

Intended applications

The technical report's abstract lists virtual reality, physical simulation, game development and interactive content creation as the intended applications ^[2]. These share a requirement that a plain image does not meet: the result has to be a space that can be entered and changed. The authors also claim state-of-the-art performance based on their own experiments ^[2].

Versions after the first release

The line moved quickly. HunyuanWorld-1.0-lite, a quantized version intended for consumer GPUs such as the 4090, followed on 15 August 2025 ^[1]. Further releases came at a steady pace: HunyuanWorld-Voyager on 2 September 2025, HunyuanWorld-1.1, also called WorldMirror, on 22 October 2025, and HunyuanWorld-1.5, also called WorldPlay, aimed at real-time world creation, on 18 December 2025 ^[1]. HY-World-2.0 was released on 16 April 2026 ^[1].

The technical report itself was first submitted to arXiv on 29 July 2025 ^[2].

Where it sits

The release made a complete world-generation pipeline available as open weights, together with named checkpoints that separate text-driven and image-driven generation from scene and sky completion ^[1]. The sequence of versions, from a quantized lite build for consumer hardware through to a real-time model, traces a steady shift from producing worlds slowly on large machines toward producing them on ordinary GPUs and, eventually, interactively ^[1].