LGM (Large Multi-View Gaussian Model)

LGM skips meshes entirely and generates 3D Gaussian primitives directly. Its asymmetric U-Net takes multi-view images — produced from a text prompt or a single image by a multi-view diffusion model — and fuses them into a set of Gaussians that render in real time. Training at 512 resolution was the key trick: earlier feed-forward models were stuck at low resolutions because of compute costs, and LGM's high-throughput backbone broke through that ceiling while keeping generation to about five seconds.

It was an ECCV 2024 Oral and became one of the foundational models for feed-forward 3D generation, with the code (MIT) and weights on Hugging Face getting heavy reuse in follow-up work like L4GM. If you need meshes rather than splats, the repo documents conversion to smooth textured meshes. The main trade-off is inherent to the approach: Gaussians are great for novel-view rendering, but mesh extraction is a post-processing step, not the native output.

• Paper-reported (arXiv 2402.05054): high-resolution 3D Gaussians from text or single-view images in ~5 seconds
• ECCV 2024 Oral; author-reported user studies and baseline comparisons; no independently verified headline numbers