But HDR bracket fusing is still a thing, no? It is just integrated in regular software. A better approach could be improving the focus stacking software, so they can correct for perspective. I think using Gaussian splats there could do that.. and eliminate artefacts too. Somebody should write a paper. I'm vary of using Ai, these are biological specimens, I don't want to add / generate detail where there were none. That would make these splats useless for any entomologist. That being sad, denoising would probably be ok(?), but buying more speedlights would be cheaper.

I think with the stacking software its the same as with the HDR-bracket-fusing software: They had their time and they have their usecases, but in gaussian splatting it'd be better not using that. I totally understand that posing the shallow-depth photos is almost impossible. would be cool, to figure out how to work around that. Regarding that noise, I have made very good experience using the AI based denoising in Adobe Lightroom. It just takes hours...

This is 1 mio, I have a 5 mio version. Unfortunatly it is actually the limit on many fronts. While the D810 does have more resolution, I need to shoot in DX, where the whole insect is in frame. The 5 mio splat actually starts to show signs of noise, I don't have enough light to shoot with lower ISO, I'd need to recharge the batteries half way through and risk a drifting specimen. Also davinci on the free version and Postshot on the cheaper version have a limit of around 3k per image. Then.. postshot cannot reliably train beyond around 6 mio splats on my machine, I'd need a graphics card with more ram or train in slices. If (when!) I improve on all these fronts, a 100 mio splat or so will need LOD to run on any lower end hardware. You are correct that focus stacking first leads to geometrically incorrect images, that is why I need to use as long a lens as possible. 90mm with 20mm extension on DX sensor is about as wide I can go, below that and the stacked images won't track anymore. You are underestimating the problem of posing the unstacked photos. They are mostly blurry, only a thin slice is in focus, that doesn't track well. Also masking the area in focus is not trivial at all, years of research and programming voodoo went into stacking software, would be a shame not to use that.

I assume that the resolution of this splat is not remotely close to the resolution of your 2D-Photos. Don't get me wrong, this kind of macro looks gorgeous, but I think it cannot live up to the quality of the captures that you've actually taken. I assume that you stack photos that have not been refocussed in lens, but that have been moved on a macrosled. By doing that, a single posed fram is a composition of multiple perspectives from different vantage-points. Something that is gemetrically incorrect. I would love to perform tests on the original images. Or I would love to see you trying to pose the images without stacking them in other software. If it is possible to create good poses from the original images (which show the positions that the sled moved to), then after exporting colmap masks could be generated with photoshop. Simply masking out the low-frequent (out-of-focus) areas.