SplatLab: Autonomous Scientific Discovery for 3DGS Research
The authors report gains of 0.18–3.78 dB over reproduced baselines, with every reported number tied to a logged training run.
SplatLab reproduces a published 3D Gaussian Splatting baseline, proposes extensions, runs training and ablations, and writes a manuscript. It ranks candidates by measured rendering quality and penalizes those that fail stress tests. A run cannot finish while any reported claim lacks supporting evidence. The authors say the process takes days of mostly autonomous compute, with code, manuscripts and experiment ledgers available on the project page.
HF Daily Papers' note
SplatLab reproduces a published 3D Gaussian Splatting baseline, proposes extensions, runs training and ablations, and writes a manuscript. It ranks candidates by measured rendering quality and penalizes those that fail stress tests. A run cannot finish while any reported claim lacks supporting evidence. The authors say the process takes days of mostly autonomous compute, with code, manuscripts and experiment ledgers available on the project page.
HF Daily Papers' note
score 5