CC BY-NC 4.0

Atrium Godiva (10M SH3, Spirula Studio)

1.6K

•172

•122.79 MB

•5mo ago

This scene is former Sandford Fleming Atrium (commonly known as "The Pit") at the University of Toronto. It was captured in January 2025, months before its renovation, with a Samsung S21 smartphone camera (~6400 frames, 1440x1080). Trained with Spirula Studio (https://github.com/harry7557558/spirula-studio, commit f3624182a39a6bab3f8b7f7f464f8f7b5c0985da; formerly spirulae-splat). This splat (10M SH3) is a follow-up of the following splats: - 6M SH2: https://superspl.at/scene/292b4172 - 8M SH2: https://superspl.at/scene/0b4360e8 Training command: `ns-train spirulae-preset-confined --data ./Atrium_Godiva/ --pipeline.datamanager.cache_images disk --pipeline.model.mcmc-cap-max 10000000 --pipeline.model.sh_degree 3 --pipeline.model.bilagrid_type loglinear --pipeline.datamanager.load_depths False --pipeline.model.alpha_reg_weight 0.0 --pipeline.model.alpha_loss_weight 0.0 --pipeline.model.alpha_loss_weight_under 0.0 --pipeline.model.num_loss_scales 1 --max-num-iterations 90000 --pipeline.model.stop-refine-at 85000 nerfstudio-data --load_thumbnails False`. Normal maps predicted by Metric3D v2 was used to supervise training. Entire training took around 6 hours on a rented remote RTX 5090. Peak VRAM usage was around 18 GiB. Special thanks @launchedpix for alignment in Metashape. Dataset download link can be found in MrNeRF & Brush Discord group. Note: SuperSplat compression seems to hurt quality quite a bit. The original PLY file can be found here: https://drive.google.com/file/d/1iSWVzLeYDaIlcAkOPp87ZAU-vDQjmsxo/view

1 Comment