TripoSR: Fast 3D Object Reconstruction from a Single Image

Tochilkin et al. release TripoSR, a FeedForward single-image 3D reconstruction model that takes the TransformerArchitecture introduced by LRM and pushes it on data, training, and engineering. The headline contribution is throughput: TripoSR returns a mesh from a single image in under half a second, compared to the roughly five-second inference of LRM, while improving quantitative and qualitative reconstruction quality on standard benchmarks.

The architecture preserves LRM's LargeReconstructionModel template — an image encoder followed by a cross-attention transformer that predicts a triplane-parameterised volumetric representation — but tightens it on several axes. The training corpus is expanded and cleaned with quality-aware filtering so that low-quality multi-view supervision is de-emphasised; new data augmentations expose the model to a wider range of object scales, orientations, and lighting conditions; and the training objective is adjusted to balance silhouette, photometric, and geometric terms in a way that the report finds more stable than the original LRM recipe. Inference is further accelerated by careful kernel scheduling and mixed-precision execution.

The technical report deliberately stays close to LRM in headline architecture; its contribution is the careful tuning that converts an academic prototype into a robust, fast, openly licensed model. TripoSR is released under MIT and has become a popular base for downstream single-image-to-3D work, including pipelines that pair it with multi-view diffusion in the style of Instant3D and pipelines that swap its implicit output for a GaussianSplatting head similar to LGM. The report is brief because the abstract claims are deliberately narrow — the value is in the released artefact and the engineering recipe rather than a new theoretical result.