Evaluation of Data Capture

Photogrammetry/Lidar scanning

The limiting factor in this approach I had so far seems mostly to be just getting usable textures.

Other issues like having enough useable photos (a second pass possibly to ensure absolute coverage), and lighting have been less critical.

I would absolutely use this technique again for complicated, non regular objects.

Retopology

I would put this mostly down to practise. I simply don't have enough hours creating UV maps and using Maya to output the best standard models for use further down the pipeline. A definite target for improvement if photoscanning is a method I use for my main project.

Learning a way to import all the LODs produced by Mudbox into Unreal might also be a good target for improvement to the workflow, instead of having to import each LOD separately.

Texturing

I don't think a full recreation of textures is the best idea for some objects. For the smaller freestanding rocks it works reasonable well, a good generic rock texture can be created with some curated details of things like moss where they would be on a real rock.

But the larger big scans already have pronounced cracks and roughness that is very time consuming to match to new textures. It's more time efficient to create a generic texture that looks like a rock face, but when compared with the original photoscans they are obvious in the lack of real world nuances that you can put into the smaller scans.

I think the ideal solution would be to able to touch-up the original photoscanned textures, with some additional layers to add detail if needed. Mixer is highly useful in how it creates a whole suite of maps that you can import easily into Unreal however.

Or alternatively use a low poly LOD and rely on the displacement maps from Mixer to create the cracks and features.

An additional issue are the texture sizes. Mixer only supports 8192px as experimental.