LGM (Large Multi-View Gaussian Model)
High-resolution 3D Gaussians from text or a single image in about 5 seconds — MIT open source.
About
LGM tackles the resolution ceiling that held back early feed-forward 3D models. Its two ideas are a multi-view Gaussian feature representation — Gaussians are a compact, renderable 3D format — and an asymmetric U-Net backbone that processes multi-view images at high throughput. The views themselves come from multi-view diffusion models, so the system works from either a text prompt or a single input image. The authors report 3D objects in about 5 seconds while pushing training resolution to 512.
The code is MIT-licensed with over 2,100 GitHub stars, pretrained weights are published on Hugging Face, and there is a Gradio demo to try. The 5-second figure is author-reported and there is no independent benchmark table, but the repo is one of the most-starred image-to-3D codebases of its era. Outputs are Gaussian splats rather than meshes, which matters for downstream use: great for rendering, less direct for game engines that want topology.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No published formal benchmarks; ~5s generation at 512 training resolution is author-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: paper (arXiv Feb 2024); MIT-licensed code with 2,100+ stars; pretrained weights on Hugging Face plus a working Gradio demo. Strengths: permissive license, high community adoption, real usable weights. Docked for: (1) speed and quality figures are author-reported; (2) outputs are Gaussian splats, not meshes, limiting engine workflows; (3) repo activity tapered after 2024. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · arXiv Feb 2024; MIT code, 2,100+ stars
- · Pretrained weights on Hugging Face; Gradio demo
- · ~5s generation at 512 resolution is author-reported
- · Outputs are Gaussian splats, not meshes