MVDream

MVDream (Multi-view Diffusion for 3D Generation) is the August 2023 ByteDance research model that attacked text-to-3D's most embarrassing failure: the Janus problem, where 2D-lifting methods grow extra faces on every side of a generated object. Instead of distilling a single-view diffusion model, MVDream jointly trains on multi-view renderings of 3D assets plus standard 2D image-text pairs, producing geometrically consistent multi-view images from a text prompt that then serve as a stable prior for Score Distillation Sampling into NeRFs and meshes.

The paper (arXiv 2308.16512) showed it matching or beating state-of-the-art 2D-lifting methods on stability while keeping 2D diffusion's generalizability — including counterfactual objects never seen in 3D data — and demonstrated DreamBooth-style few-shot personalization that preserves multi-view consistency. The honest picture: native resolution is 256×256, so fine detail needs upsampling or refinement passes, and it is a research artifact rather than a product — expect to wire it into a Threestudio-style pipeline yourself rather than getting one-click assets.