Decision Guide: If a Feature XML file is available for each part, the Hybrid preset is the first choice—it takes technical characteristics into account and reliably separates parts even when the same shape appears in many sizes. If no additional data is available, or if you have a mixed or unknown inventory, use visual_raw. For catalogs with a high proportion of sheet metal parts, discuss visual_3d and a pilot run early on.
Setting unrealistically high expectations for cluster purity—even in the best-case scenario, a large portion of the clusters will remain mixed because many elements in the dataset simply have no true siblings.
Don't expect purely visual presets to include standard designations or be categorized by size—they show the shape, not the features; that's what the hybrid presets are for.
Protection against empty clusters cannot be guaranteed across the board—it depends on the selected preset .
Promoting the three-channel preset as the best—it's an option for brass and mixed instrument catalogs, not an improvement on the hybrid preset.
Runtimes Off the top of my head—the fastest and slowest presets differ by a factor of several times, and consistently heavy sections multiply that difference even further.
A preset or a request using the old language model (gemma3) for the new image—it fails at runtime, and the error message is misleading.
Determine the memory requirements based on the old measurements, even though the new language model has not yet been evaluated.