Zero-1-to-3

Zero-1-to-3, from Columbia University and Toyota Research Institute and presented at ICCV 2023, was an early and widely adopted model for imagining how an object looks from another angle given just one photo. It fine-tuned an image-conditioned Stable Diffusion model on Objaverse renderings so that a relative camera rotation becomes part of the conditioning, and that view-aware prior went on to power single-image 3D work in threestudio, Stable-Dreamfusion, DreamGaussian and Stability AI's Stable Zero123. The repository's last commit was in September 2023, after the larger Zero123-XL checkpoint was added, and setup questions are now answered only by other users. It still runs, but the novel-view demo needs roughly 22 GB of VRAM and an older Python 3.9 environment with taming-transformers and OpenAI CLIP installed from source. Code and the released weights are MIT licensed, unlike Stable Zero123, which is non-commercial. For new work, Zero123++ produces six consistent views in a single pass, and feed-forward generators such as TRELLIS.2 or Hunyuan3D 2.1 go straight from an image to a textured mesh.