VFusion3D

VFusion3D, an ECCV 2024 paper from Meta GenAI and Oxford's Torr Vision Group, attacks the 3D data scarcity problem sideways. Instead of collecting more 3D data, it fine-tunes a pretrained video diffusion model to unlock multi-view generation, synthesizes a dataset of nearly 3 million multi-view examples, and trains a feed-forward 3D generative model on it. The result turns a single image into a 3D asset in seconds, and the authors report users preferring its outputs over 90% of the time against state-of-the-art feed-forward 3D models.

Meta released the code and weights: the majority of the project and the weights are under CC-BY-NC — usable for research, not for commercial products — with some components under Apache 2.0 and NVIDIA proprietary terms. There is a Hugging Face model card and a Gradio app for local inference. The >90% preference figure is the authors' own user study, not an independent evaluation.

Paper-reported: users preferred outputs >90% of the time vs SOTA feed-forward 3D models (author-run study)