Instant3D: Fast Text-to-3D with Sparse-View Generation and Large Reconstruction Model

Li et al. propose Instant3D, a feed-forward TextTo3D pipeline that sidesteps the slow per-prompt optimisation of ScoreDistillation methods while avoiding the low-quality outputs typical of earlier direct-prediction networks. The system splits the problem in two: first, generate a sparse set of consistent views of the desired object; second, regress a 3D RadianceField directly from those views.

The first stage fine-tunes a 2D text-to-image Diffusion model to emit four structured, view-consistent images of an object in one shot. The spatial layout of the four views is canonicalised so that downstream regression can rely on known relative camera poses. The second stage uses a TransformerArchitecture sparse-view reconstructor — a LargeReconstructionModel in the style of LRM, but conditioned on the sparse multi-view input rather than a single image — to predict the radiance field in one forward pass. This decouples the appearance/diversity problem from the geometric reconstruction problem and avoids the multi-view inconsistency ("Janus") artefacts that plague single-image optimisation methods.

End-to-end Instant3D generates a 3D asset in roughly twenty seconds, two orders of magnitude faster than per-prompt SDS systems such as DreamFusion, Magic123, or ProlificDreamer, and the diversity benefits from the diffusion sampler propagate to the 3D output. Because the multi-view generator is reusable across prompts and the reconstructor is amortised, batched generation across many prompts is straightforward, which the authors highlight as a practical advantage for asset-creation workflows. The paper is part of an emerging consensus that feed-forward reconstruction conditioned on multi-view diffusion outputs is the right factorisation for fast TextTo3D; later work (LGM, TripoSR) follows the same template, often substituting GaussianSplatting or improved transformer backbones for the radiance field predictor.