SplatLab autonomously reproduces and extends 3DGS papers
Original title: SplatLab: Autonomous Scientific Discovery for 3DGS Research
On PSNR-targeted runs, manuscripts beat reproduced baselines by 0.18 to 3.78 dB, the paper says.
SplatLab is an autonomous research system for 3D Gaussian Splatting. From a prompt, a paper URL, or no input, it reproduces a baseline, proposes an extension, trains and ablates on Mip-NeRF 360 and LLFF, and writes a manuscript whose numbers are bound to logged runs. Candidates mutate plugins from a verified backbone and are scored by a cascade-weighted Elo tournament on measured GPU PSNR, with a falsifier that subtracts Elo on failure. The authors say the cycle compresses months of graduate work to days of mostly autonomous compute, and release code, manuscripts, and ledgers.
Unconfirmed: PSNR gains and the claim of no prior nebula method are author-reported.
- Every reported number must be bound to a logged training run before termination.
- A seven-category 3DGS taxonomy wraps transferable EVOLVE-BLOCK scaffolds.
- The largest cited gain is 3.78 dB on astrophotographic nebula rendering.
- The authors say no prior 3DGS method exists for that nebula task.
From the original article
1 min read at huggingface.co
Reproducing a published 3DGS paper takes weeks of expert engineering before any extension can be proposed. Prior autonomous-research systems read papers, draft hypotheses, and write code, but are limited by short fixed templates, low executability, and weak traceability, and none has been demonstrated end-to-end in a live computer-vision subfield with intense activity.
By Seemandhar Jain, Manmohan Chandraker · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.