Quadify Ultra's ML engine is what separates it from the deterministic editions. Rather than applying a fixed algorithm to every mesh, Ultra analyses each mesh's geometry and routes it to the algorithm most likely to produce the best result.
When you click Quadify, Ultra extracts 18 geometric feature ratios from your mesh — values like triangle/quad/n-gon fractions, curvature ratio, boundary ratio, and surface area distribution. These are numerical descriptors only; no mesh shape, vertex positions, or filenames are ever read by the ML system.
The feature vector is passed to a KNN (K-Nearest Neighbours) classifier trained on the outcomes of previous operations. The classifier returns a confidence score for each algorithm. If confidence is high, Ultra routes automatically. If confidence is low — for example, on an unusual mesh type it hasn't seen before — it surfaces alternative suggestions in the panel so you can make the call.
Every operation you run is recorded locally. After approximately 30 operations, the model begins to reflect your specific mesh types and workflows. A studio that primarily processes vehicle scan data will see different routing than one working on character photogrammetry — both are correct for their context.
The ML status label in the panel shows routing when the model is actively guiding the decision, and displays confidence as a percentage.
In Settings, click Check for Updates to download aggregated model weights from all opted-in users worldwide. The download contains only statistical model data — no individual operation records, no mesh data, no personal information.
After each update, the status line shows the last update date and the total number of operations in the global database. Your local routing improves immediately after downloading.
Participation in the community model is controlled by the telemetry opt-in toggle, described in the next section.
The telemetry system is off by default. A one-time consent dialog appears on first use; you can also toggle it at any time in Settings. When enabled, the following data is sent to Supabase:
Mesh geometry, vertex positions, filenames, object names, and personal information are never transmitted. Your preference persists across Blender sessions. See the Privacy Policy for full details.
You are never locked into the ML recommendation. The algorithm selector in the panel lets you pick any algorithm manually. If you override the routing, Ultra still records the outcome — which helps the model learn from your correction.