Tencent Enhances Hunyuan3D Neural Network with Physically Based Rendering Support in Latest Update

Developers from the Chinese company Tencent have released updates for the Hunyuan3D machine learning model. The newly introduced version 2.1 now includes support for physically-based rendering, enabling the neural network to generate 3D objects that accurately reflect how light interacts with various materials. The complete code for the model, along with weights, inference capabilities, and pipelines, is fully accessible.

The Hunyuan3D project was launched by Tencent in the fall of 2024. The first iteration of this neural network was capable of creating 3D objects from textual descriptions or images, performing faster and more precisely than existing solutions by dividing the generation process into two phases. Initially, a diffusion model produced RGB images of the object from different angles, followed by the second phase, where these images were compiled into a single three-dimensional object.

In the winter of 2025, the developers released the Hunyuan3D 2.0 update. This update enhanced the model for 3D object generation and introduced Hunyuan3D-Paint, a neural network designed to create textures, increasing the accuracy and detail of the generated objects.

Now, the engineers have introduced Hunyuan3D 2.1. This update replaces the RGB-based texturing model with an advanced algorithm that employs physically-based rendering (PBR) principles. As a result, the neural network now generates photorealistic 3D objects, effectively capturing the properties of various materials.

In testing, Hunyuan3D 2.1 outperforms its competitors in both quality and speed:

The Hunyuan3D 2.1 family consists of two machine learning models:

Model

Description

Size

Link

Hunyuan3D-Shape-v2-1

Generates a 3D object from an image

3.3B

Hugging Face

Hunyuan3D-Paint-v2-1

Textures a 3D object

2B

Hugging Face

The project authors highlight that generating a 3D object requires 10 GB of video memory, while texture generation requires 21 GB, and creating both figures and textures necessitates 29 GB. These requirements are quite reasonable, allowing the model to run on consumer-grade graphics cards.

The project code is fully open-source. Developers have access to weights, inference, and pipelines, making it relatively easy to fine-tune or modify Hunyuan3D 2.1. Users can test the neural network on the Creation Engine portal.