Flythrough of the trained splat
Camera orbit through the trained Gaussian splat. Nothing here is
video of the original clip — every frame is rendered from the splat.
(8.6 MB)
Source frame vs. flythrough frame
Left: the source video. Right: the reconstruction, from a camera
angle that never existed in the source.
Measured
- Gaussians trained
- 503,133
- Frames registered
- 149 / 149 (100%)
- COLMAP 3D points
- 15,218
- Camera path
- colmap:149 cameras
- Flythrough
- 450 frames @ 30fps, 1920x1080
- Trained on
- RTX 5070 Ti, 16GB, locally
- Cost
- $0
What this is and isn't
- This is a Family Feud stage, not a Love Bird environment. It was chosen because it is the one source on this PC that reconstructs: a real Unreal editor viewport capture with genuine camera parallax.
- Every AI-generated clip tried FAILED to reconstruct, which is the expected result rather than a bug. The smoothest, best-looking one registered 2 of 156 frames. Diffusion video invents parallax per frame, so there is no consistent geometry for structure-from-motion to recover.
- The camera flies the path the capture actually took, not a 360 orbit. The capture covered an arc, and orbiting outside it would render empty space that reads as a failed reconstruction.
- Nothing in the flythrough is source video. Every frame is rendered from the trained splat.