tttLRM open-sources lightweight 3D reconstruction with test-time training

Anyone who has tinkered with 3D Gaussian Splatting knows the frustration: feed in a bunch of photos, and the rendered scene is often trapped in a haze of floaters and cloudy artifacts.

Researchers have just open-sourced tttLRM, short for Test-Time Training for Long Context and Autoregressive 3D Reconstruction. The whole model weighs in under 4GB and runs locally on consumer GPUs.

Stripping away the Gaussian fog

Traditional 3DGS pipelines frequently struggle with consistency, leaving weird smoke-like noise hanging in thin air and turning fine textures into mush.

tttLRM cleans that up dramatically. In demo reconstructions of courtyards and temples, the model captures crisp fabric weaves and intricate architectural ornaments without dissolving into visual noise.

Even notoriously difficult elements come through clearly. Thin overhead power lines and tiny printed text on posters, which almost always blur into oblivion in standard pipelines, remain sharp and legible.

The framework handles full-scale outdoor scenes just as effectively as it reconstructs isolated objects from a few handheld snapshots.

Compressing space with fast weights

The secret sauce lies in borrowing Test-Time Training from modern sequence modeling.

Instead of letting compute explode as you add more camera angles, tttLRM uses a dedicated TTT layer that operates with linear computational complexity. As images stream in, it compresses visual observations directly into a set of internal parameters called fast weights.

This forms an implicit representation in latent space, which is then decoded into clean 3D Gaussian Splats.

According to the team, the architecture also supports streaming inputs for progressive refinement as new frames arrive. The code is available on GitHub with setup instructions, and the model weights sit on Hugging Face at under 4GB.

Why it matters

We think tttLRM is a remarkably practical leap forward for everyday creators and developers who rely on reality capture.

Until now, pulling clean Gaussian Splats meant either waiting on heavy cloud rigs or spending painful hours manually cleaning up floating debris. Having an engine that generates production-friendly, noise-free geometry locally on an off-the-shelf GPU makes real-world asset capture radically more accessible.

On an architectural level, this is a huge validation for TTT outside of pure text language models. Showing that fast weights can act as an efficient spatial memory for multi-view vision proves linear-complexity recurrent layers have a bright future in 3D.

That said, it is not total magic. While thin wires and poster letters hold up impressively well, sparse viewpoints and harsh lighting changes will still challenge the underlying representations. Still, by packaging this into a lean sub-4GB download, the team gives the open-source community something genuinely useful right out of the gate.