Tor Łódź mod exposes AI limits

This is a strong hybrid-workflow use case, but not a “let the model build the track” use case. LLMs are best as copilots for extraction, scripting, and iteration; the hard geometry still comes from vision and reconstruction, not text.

• –A single 360 viewer is useful as reference, but it is a weak basis for metric geometry by itself; it helps more with layout cues, signage, barriers, and surface details than with exact widths or camber.
• –NeRF and Gaussian splatting can help when you have multiple views and decent camera poses, but sparse-view methods still struggle with accurate scale, elevation, and clean track edges without strong inputs.
• –The most realistic pipeline is: extract frames from the 360 tour, estimate depth/poses, derive a rough centerline and corner geometry, then refine in Blender or a track editor by hand.
• –LLMs can materially speed up the workflow by generating Blender scripts, organizing reference data, naming track segments, and checking consistency between reference images and spline-based geometry.
• –For sim-racing fidelity, drone, photogrammetry, or LiDAR data remains the cleanest path; AI can reduce manual labor, but it is not a replacement for ground-truth capture.