Feed-forward 3DGS
A generalized reconstruction method based on 3D Gaussian Splatting that trains a general network to output Gaussian representations in a single forward pass from sparse multi-view input, without per-scene optimization.
Timeline (last 90 days)
Existing feed-forward 3DGS approaches generally adopt a two-stage architecture that first generates the complete Gaussian representation and then encodes and compresses its attributes separately, which limits encoding efficiency.
The core advantage of feed-forward 3DGS methods is that, given sparse multi-view images, a single forward pass reconstructs the 3D scene without any per-scene optimization.
One of the structural flaws of the generate-then-compress pipeline is that the geometry and appearance encoders usually contain non-differentiable operations, so gradients cannot flow back smoothly during training, making the pipeline hard to optimize end-to-end.
All Facts (3)
Existing feed-forward 3DGS approaches generally adopt a two-stage architecture that first generates the complete Gaussian representation and then encodes and compresses its attributes separately, which limits encoding efficiency.
50%UnverifiedThe core advantage of feed-forward 3DGS methods is that, given sparse multi-view images, a single forward pass reconstructs the 3D scene without any per-scene optimization.
50%UnverifiedOne of the structural flaws of the generate-then-compress pipeline is that the geometry and appearance encoders usually contain non-differentiable operations, so gradients cannot flow back smoothly during training, making the pipeline hard to optimize end-to-end.
50%