ComfyUI Node
NeurCADRecon Train
Train NeurCADRecon to learn an SDF from a point cloud. Loss Terms (shown during training): • SDF: Distance error at surface points. Good: < 0.01 • Eikonal: Gradient magnitude should be 1 (|∇f|=1). Good: < 0.1 • Morse: Gaussian curvature regularization for sharp CAD edges. Good: < 1.0 • Total: Weighted sum. Typically starts 100-500, ends 10-50. Training takes 5-15 min on GPU for 10k iterations.
NeurCADRecon Train
- model
- point_cloud
- trained_model
- checkpoint_path
- training_log
◄num_iterations10000►
◄batch_size20000►
◄learning_rate0.0001►
◄loss_presetbalanced►
◄save_checkpointtrue►
◄checkpoint_dir►
◄log_interval500►
CategoryNeurCADRecon
Inputs (9)
| Name | Type | Default | Description |
|---|---|---|---|
| model | NEURCADRECON_MODEL | — | |
| point_cloud | TRIMESH | — | |
| num_iterationsopt | INT | 100001000–50000 | Number of training iterations. 10000 is typical for good quality. |
| batch_sizeopt | INT | 200005000–50000 | Number of points sampled per iteration. Higher uses more GPU memory. |
| learning_rateopt | FLOAT | 0.00010.000001–0.001 | Learning rate for Adam optimizer. |
| loss_presetopt | COMBO | balanced | Loss weight preset. 'sharp_edges' increases morse weight for CAD-like shapes. |
| save_checkpointopt | BOOLEAN | true | Save checkpoint after training for later use. |
| checkpoint_diropt | STRING | Directory to save checkpoint. Empty uses default location. | |
| log_intervalopt | INT | 500100–2000 | Print loss every N iterations. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| trained_model | NEURCADRECON_MODEL | — |
| checkpoint_path | STRING | — |
| training_log | STRING | — |