An out of focus image is just a mistake. The entire splat suffers severely from depth of field blur. PostShot is a small company that takes open source genius and sells it for money with obfuscated code. It is the software that suddenly becomes much faster after a FOSS project donated $1500 to somebody who can accomplish that. There is some weird timing-coincidence...

Nice! The dataset you shared contains an image in which the tomatoes are completely blurred (00169.jpeg). This seems to go against the best practices I've read about. Did you leave this image in on purpose? Is it to achieve a focus effect from certain angles? Wouldn't the result be even better without this image?

PostShot is a small company that has enabled GS experiments to a lot of people. A little bit of appreciation perhaps?

Hi Simon, thank you for sharing. Where I could learn more about this method?

Great result!

Re bilagrid: Nerfstudio splatfacto has it. Here are the parameters I use: ns-train 
--data path/to/ns-process-data/output 
--output-dir output/dir 
--pipeline.model.use-bilateral-grid=True optionally, you can pass: --pipeline.model.grid-shape Which defaults to --pipeline.model.grid-shape=16 16 8

Here is the dataset: https://www.kaggle.com/datasets/simonbethke/bowl-of-tomatoes

There is no bilagrid in brush. But LichtFeld Studio has bilagrid. It is just not stable enough for productive use. However, the results are already very promising!

Nice result! Sure PostShot is not the only solution, but features like ROI are really helpful. Since Brush is build on GSplat, do you know if it has bilateral grid implemented? The tomatoes are so smooth, I wouldn't wonder if so ;)

I really like your scan and alternative solution to get splats

amazing, looks perfect

i think I see a reflection of you on tomatoes ))

perfect tomatos

Those are tomatoes from my own garden :)

Beautiful tomatoes @simonbethke! I also agree, there’s a huge demand right now for a proper e2e processing tool in Gaussian Splatting. RealityCapture / Reality Scan is still insanely good. What really blows my mind is that even if you want a solid Gaussian splat, you still kind of need RealityCapture as a base :). Gaussian splatting actually unlocked my Tomato problem too, the typical food scanning nightmare with glossy surfaces :)

I can nearly taste them!