Pixal3D by TencentARC

Pixal3D by TencentARC

Builds a detailed 3D mesh from a single image, with pixel-aligned geometry and PBR textures. For graphics researchers and developers who run image-to-3D on their own NVIDIA GPU.

Made by TencentARC on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Pixal3D turns one image into a detailed 3D asset, saved as a GLB mesh. Instead of loosely mixing image features in, it projects pixel features back into 3D so each pixel maps to a spot in the model. The authors say this gets close to reconstruction-level detail, with PBR textures.

Who it suits: People who need a 3D model from a picture and are comfortable setting up a Python and CUDA environment.

Industry | Who uses it | How they use it | Example
Games | 3D artists | Start a model from concept art | A character picture becomes a base mesh
E-commerce | Product teams | Make a 3D view from a product photo | A shoe photo becomes a GLB file
Film and VFX | Asset teams | Block out props from reference images | A prop photo becomes a mesh
Research | Graphics researchers | Reproduce or extend a SIGGRAPH 2026 method | Run the paper branch

You can skip it if you have no NVIDIA GPU, or if you want a one-click tool. The online demo is the lighter way to try it.

Where to use it:

• A Python environment set up like TRELLIS.2, on a machine with a CUDA GPU
• The command line, through inference.py
• A local Gradio web page, started with app.py
• The Hugging Face Space demo, in a browser with no install

Use case | Who | What they get back
Single-image to 3D | Artists | A GLB mesh with PBR textures
Low-memory runs | Hobbyists | The same job with lower peak VRAM, at 1024 resolution by default
Paper reproduction | Researchers | The paper branch, built on Direct3D-S2

How to set it up:

• Python and pip, plus a CUDA GPU
• The TRELLIS.2 base environment, set up first by its own guide
• The natten package, built for your GPU architecture
• The utils3d wheel that the card names

1. Follow the TRELLIS.2 installation guide to set up the base environment.
2. Install the extra requirements.
3. Install natten, replacing xx with your CUDA architecture and worker count.
4. Install utils3d using the wheel command on the card.

pip install -r requirements.txt

NATTEN_CUDA_ARCH="xx" NATTEN_N_WORKERS=xx pip install natten==0.21.0 --no-build-isolation

How to use it: Run inference.py with an input image and an output path, and you get a GLB mesh. Add the low VRAM flag to load models on demand. The default resolution is 1536, or 1024 in low-VRAM mode.

python inference.py --image assets/images/0_img.png --output ./output.glb

python inference.py --image assets/images/0_img.png --output ./output.glb --low_vram

python app.py

Example: You run inference.py on a photo of an object with the output set to output.glb. It writes a 3D mesh you can open in a 3D viewer or game engine.

For newcomers and experts: If you are new, try the online demo first, since the local setup needs a CUDA GPU and several installs. Then run inference.py on the sample image with the low VRAM flag.

If you know the field, the main branch uses the Trellis.2 backbone for better performance, and the paper branch uses Direct3D-S2 and matches the paper's reported results. Without flash_attn, set ATTN_BACKEND=sdpa to use PyTorch's built-in attention. Resolution can be forced with the resolution flag.

One input image
Pixel features back-projected into 3D
Detailed geometry with PBR textures
GLB mesh file

Good to know:

• License: MIT
• Cost to run: your own GPU hardware, with no paid service named
• Data it reaches: the images you give it and the files it writes
• Not for: the Hugging Face Spaces requirements file, which the card says may not fit other GPU architectures

On Hugging Face, 13 November 2026

• Task: image to 3d
• Downloads in the last 30 days: 0
• Likes: 402
• License: mit
• Quantized versions: 3
• Fine-tunes: 7
• Tags: image-to-3d

Links

• Model: https://huggingface.co/TencentARC/Pixal3D
• Files: https://huggingface.co/TencentARC/Pixal3D/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download TencentARC/Pixal3D

Then follow the usage in its model card. Nothing in this kit runs by itself.