WorldFoundry
English中文

Splatt3R

Method overview: Method overview. We encode the two uncalibrated images using MASt3R’s pretrained ViT encoder and cross-attention transformers, which we freeze during training. In addition to MASt3R’s prediction head for point.

Splatt3R from University of Oxford (Active Vision Lab): 3d-reconstruction and gaussian-splatting; Verified.

Edit this pageGitHubMarkdown

Splatt3R predicts a 3D Gaussian splat scene directly from an uncalibrated image pair — zero shot, no camera parameters, no per-scene optimization. In this catalog it is fully verified: checkpoint-backed GPU inference produced a finite 344,064-Gaussian PLY from the bundled official-style views, with evidence in inference_evidence.yaml.

Method overview: Method overview. We encode the two uncalibrated images using MASt3R’s pretrained ViT encoder and cross-attention transformers, which we freeze during training. In addition to MASt3R’s prediction head for point.
Method overview: Method overview. We encode the two uncalibrated images using MASt3R’s pretrained ViT encoder and cross-attention transformers, which we freeze during training. In addition to MASt3R’s prediction head for point.

The recorded paper is Splatt3R: Zero-shot Gaussian Splatting from Uncalibrated Image Pairs (2024).

Project page: splatt3r.active.vision. Code: btsmart/splatt3r.

Build your run

Pipeline
Splatt3RPipelineworldfoundry.pipelines.component_pipelines
TaskDefault default
Environmentworldfoundry-unified-cu128
DeviceCUDA 11.3
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  splatt3r \
  --pipeline.task-profile default \
  --json

What it is for

splatt3r is the recorded reconstruction line. Start here to emit geometry from the documented views.

Per the paper (arXiv 2408.13912), Splatt3R builds on MASt3R's two-view geometry backbone and adds a Gaussian-parameter head. WorldFoundry keeps the runtime under base_models/three_dimensions/general_3d/splatt3r/splatt3r_runtime, importing MASt3R/DUSt3R from their canonical sibling paths; the packaged runtime is inference-only (training loaders and demo code are pruned). It is published by University of Oxford (Active Vision Lab).

Catalog tasks include 3d-reconstruction and gaussian-splatting. Catalog aliases include splattr and splatt3r-official.

Runtime contract

Optional input_path and interactions; artifacts splatt3r.ply.

CLI task profiles: default. WorldFoundry binds Splatt3RPipeline.

Family status: Verified.

Splatt3RSynthesis loads the in-tree runtime in-process and emits a PLY Gaussian-splat artifact. Checkpoints stay external assets (brandonsmart/splatt3r_v1.0) — stage them before running.

Runtime code is vendored in-tree under worldfoundry/base_models/three_dimensions/general_3d/splatt3r/splatt3r_runtime; checkpoints remain external assets.

Running in WorldFoundry

The catalog launch is worldfoundry-eval run splatt3r --pipeline.task-profile default. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 11.3).

Recorded weight repositories include brandonsmart/splatt3r_v1.0. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh splatt3r. Add --download to fetch missing public checkpoints.

Status: Verified.