Open-source text and image to 3D asset generation with PBR materials
Tencent Hunyuan 3D generates 3D assets from text or images using diffusion models, with PBR material output. Code, weights and training pipelines are open-sourced, so it suits both hands-on creators and developers who want to fine-tune.
Tencent Hunyuan 3D is a diffusion-based 3D generation system from Tencent that produces 3D assets from text prompts or images. The site presents several model generations side by side (1.0, 2.0, 2.1 and Omni), with the 2.1 release positioned as the current showcase. Its pipeline combines text and image encoders, a diffusion model and a 3D decoder, which supports multi-view generation, reconstruction and single-view generation. The standout claim is physically-based rendering (PBR) texture synthesis. Instead of the plain RGB textures used by earlier versions, the 2.1 pipeline simulates physics-grounded parameters such as diffuse reflection, metallic properties and normal maps. That is meant to give materials like leather and bronze believable light behavior, which matters for game assets and industrial design where flat color textures look dated under real-time lighting. Tencent also stresses openness. Data processing pipelines, training and inference code, model weights and architecture are described as released, with links to GitHub, Arxiv and Hugging Face. Developers and researchers can therefore fine-tune the model for downstream tasks, reproduce baselines or self-host it, rather than relying only on a hosted service. A "Try It Now" button leads to the hosted web experience for people who want results without setting anything up. In practice, the audience splits in two. Game developers, 3D artists and product designers can use the hosted interface to get a starting mesh with materials from a sentence or a reference picture. Machine learning teams and academics can pull the weights and code to build on. Compared with closed commercial generators, the draw is the open release and the PBR output; the trade-off is that the homepage offers little about pricing, usage limits or support, and the interface is oriented toward a Chinese-market audience, though an English toggle is present. It suits users who value control and transparency over a polished, fully documented commercial workflow.
Text-to-3D and image-to-3D generation, multi-view generation, reconstruction and single-view generation. PBR material synthesis (diffuse, metallic, normal maps). Open-sourced weights, training and inference code, and data pipelines for fine-tuning.
