TTTLRM project brings long-context 3D reconstruction to consumer GPUs

A new research project called TTTLRM promises higher-quality 3D scene reconstruction from simple photo dumps. What makes it stand is it doesn’t require datacenter-grade hardware.

The model turns multi-view photos into detailed 3D Gaussian splats. In addition, it’s sized under 4 GB, which is enough to run on consumer GPUs.

TTTLRM stands for “test-time training for long-context autoregressive 3D reconstruction”. It uses fast weights that update during inference to better fit each scene.

In benchmarks, it beats existing 3DGS-based methods on detail and consistency while staying efficient enough for creators, AR/VR developers, and hobbyists. The team has released both code and weights on GitHub and Hugging Face, making it easy for users to try the model on their own photo collections.