Best by far in a minute! Kinda fun trying to find things that can render this... Supersplat is a no go, Brush has slightly better speed but still slow. The benchmarks must commence!
iliesaya -
I aligned the cameras with RealityScan and processed using Postshot.
If your model is blurry after the appropriate number of training steps, either something is wrong with your images or the alignment. Can you look at the point cloud before you start processing the model?
I would be careful using focus stacking due to lens breathing or digital artefacts.
fantastic result !! would you mind sharing the step after capturing image ? camera alignment with Metashape/ Colmap/ RealityCapture ? Gaussian processing with PostShot / LichField ? I did similar scan with 47Mp pixel camera, 800 Focus Stacking image, but i am no where close to your result, its always very blurry .
Thanks slimbuck. Yes, using larger sensors does mean more challenges with DOF, but it also comes with higher-quality images and point clouds for alignment.
I used around f/29, which is quite high and has diffraction, which makes it less sharp overall, but I think it is better than using a wider aperture. There is probably a happy medium.
I used lens tilt to adjust the focus plane, that's when the sensor stays level but the lens tilts up or down.
It's a manual focus lens, so focus stacking would be very tedious. If it were electronic, it could be automatically captured by the camera provided lens breathing isnt an issue.
It's quite a large model, so yes, some LOD settings or optimising is needed for real use. NIRA.app can be used to stream very large mesh and 3DGS models to low power/data clients, although this is not really 'consumer' focused and more focused on AECO.
Incredible capture!
I'm a little surprised medium format sensor works so well on an object of this size. I would have expected the relatively small DOF to work against you?
Do you mind sharing what aperture and/or focus stacking you did to achieve this?
It would awesome to see a streamed version of this model, which could then be accessed on any device (including low-end mobile devices).
If you have multiple versions of this model at various LODs (or are able to generate them) and are at all interested, I can help get this published.