Single-image 3D reconstruction escapes the paywall

What it does

OpenLRM is an open-source reproduction of the LRM paper: feed it a single image—ideally RGBA or a masked RGB with a clean background—and it emits a 3D mesh and a rendered video. Under the hood it relies on a DINOv2 encoder, with pretrained weights ranging from small to large. The repository also includes training scripts and a Blender pipeline for rendering Objaverse images, so you can roll your own dataset if the pretrained mix isn’t enough.

The interesting bit

The authors release two weight families: “obj” variants trained only on Objaverse, and “mix” variants trained on Objaverse plus MVImgNet, with the latter acting as the better general-purpose default. The whole pipeline is HuggingFace-native, offering automatic weight downloads and a web demo, but the weights are locked under a CC BY-NC 4.0 license—fine for research, useless for products.

Key highlights

• Single-image inference to mesh or video via DINOv2-based reconstruction
• Six pretrained checkpoints (small / base / large, each in obj-only or mixed-data flavors)
• Memory-efficient attention via xFormers inside the DINOv2 encoder
• Training code and Blender data-rendering scripts for Objaverse
• HuggingFace demo and model-hub integration for quick experiments

Caveats

• Model weights are CC BY-NC 4.0 and certain components carry an NVIDIA proprietary license, so commercial use is off the table.
• Default training assumes 8 GPUs with bf16 mixed precision, and inference can run out of CUDA memory unless you shrink frame_size.
• PyTorch >=2.1 is recommended, and xFormers must be installed separately to unlock efficient attention in the DINOv2 encoder.

Verdict

Researchers and hobbyists building 3D generative pipelines should grab this; commercial teams and anyone without a decent GPU should look elsewhere.