DreamGaussian

open_source

DreamGaussian generates high-quality textured 3D meshes from a single image or text prompt in about 2 minutes using Generative Gaussian Splatting. Open-source, ICLR 2024.

AI Models & Infrastructure

AI Image Generators

3D Modeling Tools

About

DreamGaussian is a novel 3D content creation framework introduced at ICLR 2024 (Oral) by researchers from Peking University, Nanyang Technological University, and Baidu. It addresses a critical bottleneck in modern 3D generation pipelines: slow per-sample optimization. Traditional score distillation sampling (SDS)-based approaches can take 20+ minutes per object; DreamGaussian collapses this to roughly 2 minutes without sacrificing quality. The core innovation is a Generative 3D Gaussian Splatting model paired with a two-stage pipeline. In Stage 1, the model progressively densifies 3D Gaussians, which converges far faster than occupancy pruning used in Neural Radiance Fields. In Stage 2, an efficient algorithm converts the Gaussian representation into a textured mesh in UV space and applies a fine-tuning pass to sharpen surface details, making outputs immediately usable in downstream applications such as animation via tools like Mixamo. DreamGaussian supports three workflows: image-to-3D (including images with non-zero elevation angles), text-to-3D, and a combined text-to-image-to-3D pipeline. The resulting meshes are clean, animatable, and exportable in standard formats. The full codebase is publicly available, making DreamGaussian accessible to researchers, 3D artists, game developers, and anyone building generative 3D pipelines on consumer-grade GPUs (demonstrated on an NVIDIA RTX 3070 8 GB).

Key Features

Image-to-3D in 2 Minutes: Converts a single-view image (including non-zero elevation angles) into a fully textured 3D mesh approximately 10× faster than SDS-based methods.
Generative 3D Gaussian Splatting: Uses progressive densification of 3D Gaussians instead of occupancy pruning, enabling significantly faster convergence for generative 3D tasks.
Mesh Extraction & UV Texture Refinement: Automatically converts Gaussian representations into clean, export-ready textured meshes and applies a fine-tuning stage in UV space to sharpen surface details.
Text-to-3D & Text-to-Image-to-3D Pipelines: Supports text-driven 3D generation by integrating with text-to-image models, enabling fully text-guided 3D asset creation workflows.
Animatable Mesh Output: Exported meshes are compatible with standard animation tools such as Mixamo, making them ready for games, VR, and cinematic pipelines.

Use Cases

Rapidly prototyping 3D game assets from concept art or reference images without manual modeling.
Generating animatable 3D characters or objects for VR/AR experiences from text descriptions.
Accelerating 3D content pipelines for digital media, film, and visual effects by converting 2D artwork into 3D meshes.
Academic and ML research into generative 3D representations, Gaussian Splatting, and novel-view synthesis.
Creating 3D product visualizations or e-commerce assets from single product photographs.

Pros

Exceptional Speed: Completes 3D generation in ~2 minutes on a consumer GPU (RTX 3070 8 GB), roughly 10× faster than competing optimization-based approaches.
Fully Open-Source: Code and models are publicly released, enabling researchers and developers to reproduce results, fine-tune, or integrate into their own pipelines.
Production-Ready Mesh Output: Outputs textured meshes in standard formats that can be directly imported into 3D software and animation tools without heavy post-processing.

Cons

Requires Technical Setup: As a research codebase, DreamGaussian requires a Python/CUDA environment and GPU hardware — there is no polished GUI or hosted web app.
Single-View Limitations: Generation quality depends heavily on the quality and pose of the input image; complex geometries or occluded regions can produce artifacts.
No Commercial Support: Being an academic project, there is no dedicated support team, SLA, or regularly maintained cloud service for production deployments.

Frequently Asked Questions

DreamGaussian is an open-source 3D content generation framework presented at ICLR 2024. It uses Generative 3D Gaussian Splatting to create textured 3D meshes from a single image or text prompt in approximately 2 minutes.

DreamGaussian is approximately 10× faster than existing score distillation sampling (SDS)-based 3D generation methods, producing results in ~2 minutes versus the 20+ minutes typically required.

It supports three workflows: image-to-3D (single-view images, including non-zero elevation angles), text-to-3D, and a combined text-to-image-to-3D pipeline.

The paper demonstrates results on an NVIDIA RTX 3070 with 8 GB VRAM. A CUDA-capable GPU is required; performance will scale with GPU capability.

Yes. DreamGaussian exports standard textured meshes that are compatible with animation platforms like Mixamo and can be imported into game engines or 3D software such as Blender and Unity.