BAE · MULTI-VIEW RECONSTRUCTION

See what camera alignment changes.

Turn a video or photos into 3D. Compare Depth Anything 3 with BAE-refined camera poses, side by side.

0.1 60
Image Processing Method

Official DA3 setting: low_res for more images.
Select First Frame

Empty uses lexicographic filename order.
Rendering trajectory for 3DGS viewpoints
Video quality for 3DGS rendered outputs
0 100
1000 100000
BAE refinement stages

Official DA3 input and inference settings. Gaussian inference is enabled for this comparison. No frame-count limit or additional input resizing.

Also refines focal calibration. Changing this setting reuses the cached DA3 prediction.

Turn on for photos or sparse video where neighboring frames share little view. Otherwise BAE may skew the cameras.

Upload views of a mostly static scene with overlap and camera movement. Prepare the frames before reconstruction.

How to compare: both sides share depth predictions. BAE changes the cameras; optional self-calibration also changes the intrinsics. A photometric check is a consistency signal, not ground-truth accuracy.
Your paired point clouds will appear here.

Built with BAE, Depth Anything 3, and Spark. Upload only content you may use. Session inputs and results are temporary, with a one-hour lifetime. They are not committed to the Space. DA3 model: CC BY-NC 4.0.