InfiniSplat v2 turns multi-view photos into explorable 3D worlds

InfiniSplat v2 is a freshly open-sourced feed-forward 3D reconstruction model that turns multi-view photos or continuous video into interactive 3D scenes you can walk through.

Instead of making you wait through hours of per-scene optimization like traditional 3D Gaussian Splatting, it uses a feed-forward pipeline. Feed it images, and it instantly builds an explorable 3D environment.

What it can do

You can feed the model casual snapshots taken from different viewpoints, and it will reconstruct a coherent 3D object with fine geometric fidelity.

It also scales up to massive environments. In published demos, the team fed it a driving sequence with over 1,000 frames, and it reconstructed the entire roadway and surroundings with smooth camera navigation.

Most importantly, it solves the floating artifact problem. While older feed-forward models tend to break apart or hallucinate blurry clouds when you turn the camera, InfiniSplat v2 keeps surfaces clean and crisp.

Under the hood

Handling wide viewpoint changes has always been the Achilles' heel of feed-forward Gaussian methods. InfiniSplat v2 tackles this by rethinking both the architecture and training pipeline.

The core trick is decoupling. The model splits geometry estimation from Gaussian attribute prediction, relying on a frozen geometry branch so the network doesn't conflate physical shape with surface lighting.

During training, the researchers align predicted Gaussians with ground-truth camera poses and enforce large-angle target view supervision.

They also discovered that learning Gaussian opacity purely through image reconstruction supervision naturally filters out geometric mismatches. Basically, the model learns to make erroneous floaters completely invisible.

Across six benchmark datasets, InfiniSplat v2 set new state-of-the-art marks for feed-forward outputs. In pose-conditioned evaluations, mean perceptual error metrics LPIPS and DISTS dropped by 25.8% and 32.9% respectively compared to the strongest baselines.

How to run it

Both code and pre-trained model weights are already hosted on the project page and GitHub. The complete model weights take up around 6.6GB.

That compact footprint means it runs comfortably on modern consumer GPUs. The repository includes clear instructions for local installation, from dependency setup to inference commands on custom image sets.

Why it matters

We see InfiniSplat v2 as a big practical win for 3D creators, roboticists, and game developers. Moving from painful per-scene optimization to instant feed-forward inference cuts the turnaround time of world-building from hours down to seconds.

While feed-forward outputs still won't beat hours of offline training on extreme edge cases, lowering the hardware barrier to a 6.6GB model on a consumer graphics card makes neural 3D scene capture far more accessible. It is definitely worth cloning to test on your own footage.