threestudio-project/threestudio

Threestudio is a research framework for generating 3D objects and scenes from text descriptions or images. You type a prompt like "a campfire with glowing embers" and the system produces a three-dimensional object you can view from any angle. It can also take a single photograph as input and reconstruct a 3D version of what is shown. The output can be viewed, exported, or used in downstream 3D applications. The core idea behind the technology is to take AI models that generate 2D images from text, which have become very capable in recent years, and use them to guide the construction of a 3D representation. The framework does not train a new AI model from scratch for each prompt. Instead, it runs an optimization process that keeps adjusting a 3D scene until 2D views of that scene look like what the image-generation model would produce for the given prompt. This approach is called score distillation and it was pioneered in academic papers like DreamFusion. Threestudio itself is a unified codebase that implements many different methods from published research papers, including DreamFusion, Magic3D, ProlificDreamer, Zero-1-to-3, and others. Researchers and practitioners can run and compare these methods from one shared code base without maintaining separate repositories for each paper. A community extension system lets contributors add new methods without modifying the core code. Using it requires a GPU, Python, and a set of dependencies listed in the installation instructions. There is also a Google Colab notebook for trying it without a local GPU, and a Gradio web demo hosted on Hugging Face for the most basic text-to-3D use. Output quality and generation time vary by method, with some approaches taking minutes and others taking much longer depending on hardware. The project is aimed at researchers working on 3D generation and at technically experienced practitioners who want to experiment with state-of-the-art methods. The full README is longer than what was shown.