3D Dataset Generation on SimJEB v1
15,360Total 3D SamplesDeepJEB++
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DeepJEB++Seed data=380 (from SimJEB), Simulation labels=FEA (automated BC + CAE), Generation method=Foundation model transfer, 2D pretraining=Stable Diffusion (1B+), 3D pretraining=TRELLIS (500K+ shapes), Quality filtering=VLM-based classifier2026.06 | 15,360 | 40 | |
| DeepJEBSeed data=263 (from SimJEB), Simulation labels=FEA (automated), Generation method=DeepSDF interpolation, 2D pretraining=None, 3D pretraining=None, Quality filtering=Min Jacobian + IQR2026.06 | 2,138 | 5.6 | |
| SimJEBSeed data=—, Simulation labels=FEA (manual setup), Generation method=Hand-designed, 2D pretraining=None, 3D pretraining=None, Quality filtering=Manual + IQR2026.06 | 381 | 1 |