CRM (Convolutional Reconstruction Model)
Single image to textured 3D mesh in about 10 seconds — open MIT feed-forward reconstruction.
About
CRM, from Tsinghua's machine learning group and presented at ECCV 2024, is a feed-forward image-to-3D model with a refreshingly practical pitch: one input image, one textured mesh out, in roughly 10 seconds. It generates six canonical views plus canonical coordinate maps from the input, then reconstructs geometry and textures with a convolutional architecture. The official implementation produces proper UV-textured OBJ meshes (the lighter Hugging Face demo uses vertex colors and is faster but less detailed).
The code is MIT-licensed with around 700 GitHub stars, pretrained weights are on Hugging Face, and there is both an official Gradio demo and a community Replicate deployment for trying it without a GPU. The 10-second figure and the quality are author-reported; there is no formal benchmark table. Setup is fiddly (nvdiffrast, xformers, kaolin), which is worth knowing before promising it to anyone without a CUDA machine.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No published formal benchmarks; ~10s inference and quality are 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: ECCV 2024 paper from Tsinghua; MIT-licensed code (~700 stars) with released weights; official HF demo plus community Replicate deployment. Strengths: permissive license, genuinely fast inference, try-before-you-install demos. Docked for: (1) no formal benchmarks, all figures author-reported; (2) heavyweight install (nvdiffrast, kaolin, xformers); (3) no official API or maintenance cadence. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ECCV 2024 paper (Tsinghua THU-ML)
- · MIT code, ~700 stars; weights on Hugging Face
- · Official Gradio demo + community Replicate demo
- · ~10s inference is author-reported; no formal benchmarks