OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Dit 3D

DiT-3D is an open-source Python research model and codebase for generating 3D shapes with a diffusion transformer that denoises voxelized point clouds.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Dit 3D — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use DiT-3D when you need research code for training or testing a diffusion-transformer model that generates 3D shapes from voxelized point clouds.

Who it is for

  • Researchers studying diffusion transformers and 3D shape generation.
  • Developers training or testing point-cloud generation models with ShapeNet data.

Documented uses

  • Train a configurable DiT-3D model on ShapeNet point-cloud data.
  • Test a saved checkpoint and visualize generated chair point clouds.
  • Compare model configurations using the repository's evaluation workflow.
Before you choose

Requirements & limits

The code is in a public GitHub repository under the Apache-2.0 license.

  • The README specifies an older, pinned environment including Python 3.6, PyTorch 1.7.1 and CUDA Toolkit 11.0.
  • Training and testing examples require local ShapeNet data and configured paths.
  • The linked pretrained model is Small with patch size 4; the paper's main comparison is described as using XL models.
  • The repository lists separate metric dependencies, including PyTorchEMD, PointNet++ and GPU kNN.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Train a chair model with the documented configuration

Prepare and run the README's example training configuration for DiT-3D-S/4 on ShapeNet chair point clouds.

Before you start

  • Python 3.6, PyTorch 1.7.1, torchvision 0.8.2 and CUDA Toolkit 11.0, plus the other listed dependencies.
  • ShapeNetCore.v2.PC15k point-cloud data.
  • The repository's training script and its documented metric dependencies.

Steps

  1. Install the repository dependencies with the command below, and install the separately listed PyTorchEMD, PointNet++ and GPU kNN dependencies as needed.

    Documented command
    pip install -r requirements.txt

  2. Obtain ShapeNet point-cloud data using the resource linked from the README.

  3. Run the documented train.py example with category chair, model type DiT-S/4, voxel size 32, batch size 16 and the desired data and experiment paths.

Expected outcome

The documented configuration trains a Small DiT-3D model with patch dimension 4 on the chair category at voxel size 32.

Guide limitations

  • The README's example assumes local data and experiment paths that must be supplied by the user.
  • The guide describes the documented setup and command, not a verified execution or a guarantee of training results.
Documentation 1 ↗
Source-based practical guide

Test and visualize a chair checkpoint

Use the README's testing example to evaluate and visualize a DiT-3D-S/4 chair checkpoint at voxel size 32.

Before you start

  • The documented Python and PyTorch environment and repository dependencies.
  • ShapeNet point-cloud data.
  • A DiT-3D checkpoint file.

Steps

  1. Install the repository dependencies using the README's requirements instructions.

  2. Obtain ShapeNet point-cloud data and a checkpoint; the README links to pretrained models and identifies the linked model as Small with patch size 4.

  3. Run test.py with category chair, num_classes set to 1, model type DiT-S/4, voxel size 32, and paths for the data and model.

Expected outcome

The documented testing workflow evaluates the chair model and supports visualization of its generated point clouds.

Guide limitations

  • The example expects local data and checkpoint paths.
  • The README says the cited pretrained checkpoint is Small with patch size 4, while the paper's main comparison uses XL models.
  • The README names Mitsuba for point-cloud rendering; it does not establish that every testing setup includes it.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
16 October 2026
Original website title
DiT-3D/DiT-3D · Official homepage title · retained in its source language
How we review sourcesSuggest a correction