27/05/2026
For the first detailed comparison, I want to start with the baseline: photogrammetry.
This object was captured with the Plinth iOS app using auto capture on a turntable.
133 photos
15 minute capture
Full processing time: 7:45
Post-adjustment in Plinth to tune the PBR materials
What photogrammetry gets right is that the result is tied to observed image data.
The geometry is reconstructed from many captured views of the actual object. Surface detail, proportions, visible damage, material variation, and object-specific irregularities are grounded in photographs of the real object.
That does not mean photogrammetry is perfect.
It still depends on capture quality, lighting, coverage, surface properties, and enough overlap between images. Reflective, translucent, very dark, or low-texture materials can still be difficult. Post-processing and material tuning also matter.
But for cultural heritage, this baseline is important because the workflow is evidence-based.
We can say what was captured.
We can inspect the source image set.
We can understand where the geometry came from.
We can distinguish reconstruction from adjustment.
That is the reference point I want to compare against when looking at AI-generated 3D outputs from TRELLIS.2 and Apple LiTo.
The question is not simply whether the result looks good.
It is whether the result can be traced back to the object.
Object from Museum Beelden aan Zee.
View the model:
https://plinth.it/model-viewer/s/zpABUf7b