SplatLab writes its own 3D Gaussian Splatting papers and beats baselines by up to +3.78 dB

Reproducing a published 3D Gaussian Splatting paper takes an experienced engineer weeks of work, and only after that can they propose their own extension. SplatLab tries to remove this bottleneck entirely. The system accepts a research query, a link to a target paper, or nothing at all. It then reproduces the baseline, proposes an extension, trains and ablates it on Mip-NeRF 360 and LLFF, and writes a manuscript. On tasks with a PSNR target metric, its papers beat the reproduced targets by margins from +0.18 to +3.78 dB. A cycle that takes a graduate student months shrinks to days of mostly autonomous compute.

Why earlier "AI scientists" stalled

Systems that read papers, formulate hypotheses, and write code already exist. The SplatLab authors point to three weaknesses: short, fixed templates, low executability of the generated code, and poor traceability of results. In addition, no such system has been demonstrated as an end-to-end loop in a fast-moving, highly competitive area of computer vision. 3DGS is exactly that kind of field: papers arrive in a steady stream, and differences between methods are measured in tenths of a decibel.

Five design decisions inside the pipeline

First: candidates mutate plugins on top of a verified implementation that already matches published baselines, rather than code written from scratch. This restricts the search space to working variants.

Second: selection runs through a tournament with a cascade-weighted Elo rating across eight "islands." Candidates are scored by GPU-measured PSNR, and feedback arrives as a textual gradient: a written analysis of what worked and what did not.

Third: a falsifier that runs continuously deducts Elo points if a candidate fails stress tests. Robustness under scrutiny becomes part of selection rather than a separate stage after it.

Fourth: every number in the manuscript is tied to a logged training run. The system cannot finish while even one claim remains unsupported.

Fifth: a seven-category 3DGS taxonomy. EVOLVE-BLOCK markers are wrapped in domain scaffolds that transfer between 3DGS repositories.

Nebulae as a showcase case

The most notable result is +3.78 dB on rendering astrophotographic nebulae. According to the authors, no ready-made 3DGS method existed for this task. In other words, the system did not improve a known recipe; it built one for a little-studied domain. The lower end of the range, +0.18 dB, looks more modest, but in a community that competes over tenths of a decibel, even that gain counts as significant.

Key system parameters:

• Input: a topic, a target paper URL, or an empty query
• Benchmarks: Mip-NeRF 360 and LLFF
• Selection: Elo tournament on 8 islands by GPU PSNR
• Safeguard: a falsifier that deducts Elo on failure
• Traceability: every number tied to a run log
• PSNR gain: from +0.18 to +3.78 dB
• Cycle time: days instead of months of graduate-student work

What you can use right now

The authors have posted the code, all manuscripts, experiment logs, and the cross-run paper store on the project page. For labs, this is ready-made material for verification: you can take a specific manuscript and trace the chain from a number in a table back to the training log. That kind of audit is most often missing from human-written papers.

For teams that use 3DGS in products, the practical takeaway is different. The verified baseline replica and the tournament search over plugins can serve as an automated test bench for finding improvements on your own data, whether that is indoor scans, drone footage, or scientific imagery. The rule "no log, no claim" is worth adopting outside 3DGS as well: it addresses the main risk of automatically written research, where polished text diverges from what the runs actually show.

The question of limits remains open: PSNR is measurable and cheap, while many areas of computer vision lack such a clear metric. But SplatLab shows that in a subfield with clear measurement and fast experiments, an autonomous loop from reading a paper to writing a manuscript already works.