A rover or lander needs to see rocks and hazards on the Moon — but there's almost no labeled real lunar imagery. So I made the data: a domain-randomized synthetic lunar world in OpenUSD, rendered with NVIDIA Omniverse Replicator, and ran the whole pipeline — generate, train a SegFormer hazard segmenter, evaluate, and render — on a single NVIDIA DGX Spark.
Then I made the rocks photoreal. The synthetic benchmark went up (rock-IoU 0.815 → 0.852). But the honest result is the surprise: the more realistic model transfers worse to real Apollo photos — it floods ~83% of real pixels as rock — because photoreal rocks collapsed the rock-vs-regolith boundary toward “rough grey texture = rock,” and real regolith is exactly that.
The lesson: higher fidelity and higher synthetic scores don't buy better real-world transfer. Synthetic metrics can mislead — you have to test on real.