AGG: Amortized Generative 3D Gaussians for Single Image to 3D

Xu et al. propose AGG, an amortised ImageTo3D model that produces a 3D GaussianSplatting representation from a single input image in a single forward pass. Existing Gaussian-based 3D generators of the period — for example DreamGaussian — relied on optimisation-based pipelines built around ScoreDistillation, paying a per-instance compute cost on the order of minutes. AGG eliminates this loop by training a neural network that maps directly from an image to the parameters of a Gaussian set.

The architecture decomposes the generation problem along two axes. First, AGG separates the prediction of Gaussian locations from the prediction of the remaining appearance attributes (opacity, colour via spherical harmonics, scale, rotation) and lets the two streams interact through a hybrid intermediate representation. This decomposition stabilises optimisation, since locating primitives near the true surface is a much better-conditioned problem than jointly predicting all attributes at once. Second, the pipeline is cascaded across resolution: a coarse 3D Gaussian representation is produced first, and a downstream 3D Gaussian super-resolution module then upsamples primitive count and refines local detail.

In comparisons against optimisation-based Gaussian generators and against amortised models built on alternative 3D representations such as NeRF, AGG produces competitive quality while running several orders of magnitude faster — the per-image generation cost drops from minutes to a fraction of a second. AGG sits in a broader family of feed-forward Gaussian predictors that includes LGM, MVSplat, and pixelSplat; together they trace an arc from per-prompt optimisation toward learned amortised inference for 3D content creation.