Gaussian Splat - LichFeld Studio Test Results

1.3K

I did a series of tests in LichtFeld Studio to see how different resolutions and total splats counts affected the quality and processing speed of models. My PC has a 3080 FE (10GB VRAM) Process: I took 98 photos (24MP) with my DSLR camera of a vent at the side of my house, aligned them within RealityScan and then exported 3 different and separate COLMAPS from RealityScan, with different levels of downscaling of the images since I was getting crashes in LichtFelt Studio at higher resolutions. Lowest Quality = 0.5MP (shockingly low at 842 x 594) Medium Quality = 1MP (1200 x 833) Higher Quality = 10MP (3814 x 2622) I then used each data set to create separate 3DGS with a total gaussian splat count of 10k, 500k, 5M to compare how they all looked and performed. I then exported each dataset directly from LFS as SOG files (hope that wasn’t a mistake going straight for the most compressed format…) and then brought them into SuperSplat where I aligned them all in order in the one scene, tidied up the splats a little, annotated them etc. Top row is all the 0.5MP images, with increasing splat count going rom left to right. Next row is 1MP, and bottom row is 10MP. Check out the annotations! The only model I could not generate on my GPU was the 10MP images aiming for a 5M gaussian splat. LFS kept crashing on me. Notes: - I kept pretty much all the LFS settings default, other than Max Gaussians. - For some of these models processing, I was using my browser and messing about with SuperSplat, which may have skewed my VRAM usage as when that was open the number on LFS would go up, so that’s not an exact science. - I’m currently trying to reprocess the 10MP images for 5M splats without using the computer to see how it goes. It’s currently maxing out my VRAM at 10.7GB, processing super slowly (4 iterations a second) with an ETA of about 2 hours. My thoughts: - My biggest shock was how good the 0.5MP images actually worked. It was trained so fast and looks so good. Genuinely shocked! From 24MP to 0.5MP and still looking so good. - I found that once the targeted number of splats was achieved, the rest of the iterations didn’t really seem to make all that much difference. - More splats is not necessarily better. I find that having more splats resulted in way way more floaters and “noisier” splats. Not sure if it’s a SOG compression issue, but the 5M splats that processed seem arguably worse than the lower count ones. - 10k gaussians were basically unusable, but interesting to see nonetheless. Interested in hearing peoples thoughts on this as I’m just dipping my toes back into Gaussian Splats!

Suggested

2 Comments

ijenko5 months ago

oh wait... got confuddled. Postshot typically defaults to 3m splats. and 35-45k Training steps.

ijenko5 months ago

Nice testing! Am about to jump into Lichfield so this is a useful test for me also shocked about how good the 0.5mp ones looked. #2 isnt far off from being the best result! Would be interesting to see how this translates to bigger scenes where the photos wont have so much overlapping information. maybe add some more steps between 10k and 100k. Postshot usually defaults to about 30-40k splats which often results in good results. You can use postshot in free mode to get an idea how its processing times and quality stack up - you just wont be able to export the results, but will give you a good indication how it compares to lichfield .

Loading comments...