---
pretty_name: PowerSim checkpoints and simulation assets
tags:
- physics-simulation
- powerfoam
- material-point-method
---
# data/ — released assets (not in git)

| Folder | Contents | Used by |
|---|---|---|
| `powerfoam_ckpt/<name>/` | `model.pt`, `config.yaml`, `cameras_train.pt`, `cameras_test.pt` | every demo (`--checkpoint-config`) |
| `mat_prop_ckpt/bonsai_tree/` | `material_field.pt` (per-primitive E, ν recovered by material estimation, paper Fig. 1c), `triplane_fields.pt` (optimiser state), `selection.pt`, `sim_config.json` | `MATERIAL_FIELD=... run_garden_bonsai_poke.sh` |
| `selections/<name>/` | boolean primitive masks (`*_mask.pt`), per-scene material fields and other sim inputs | the run script of that scene |

## foam_evolve bundle

| Checkpoint | Size | Scene |
|---|---|---|
| `garden_v1_bonsai_foreground_edit` | 2.4 GB | MipNeRF-360 garden with the bonsai composited in place of the vase |
| `bread_roll@ec2ced65` | 0.35 GB | bread roll (Blender render data) |
| `ficus@48d229db` | 0.22 GB | NeRF-synthetic ficus |
| `pillow2sofa_gs_fewer_points` | 1.1 GB | PhysGaussian pillow2sofa, distilled to PowerFoam |
| `coke_can_mesh_fewer_points` | 0.97 GB | coke can (mesh render) |
| `wolf_gs_filled` | 1.4 GB | PhysGaussian wolf, distilled and interior-filled |
| `telephone_sfm@2ed8d7fa` | 1.9 GB | captured telephone (non-pinhole COLMAP camera) |
| `mic@84daf179` | 0.27 GB | NeRF-synthetic mic |

`cameras_<split>.pt` (written by `scripts/prep/export_cameras.py`) replaces the training images:
image size, camera-to-world poses, the eye/right/up vectors of each view and one shared set of
camera-space ray directions, from which each view's ray map is rebuilt exactly as PowerFoam's
dataset loader does.

## material_estimation bundle

| Path | Content |
|---|---|
| `powerfoam_ckpt/carnations_sfm@ae005aa7/` | the carnation checkpoint (2.1 GB) + cameras |
| `selections/carnations_sfm@ae005aa7/flower_mask_dilated20.pt` | the flower primitives (195 299 of 1.2 M) |
| `selections/carnations_sfm@ae005aa7/interior_fill_cache_flower_mask_dilated20.pt` | interior particle filling of that selection (regenerable with `powersim.material_estimation.filling.precompute_interior_fill`) |
| `gt_video/carnation/carnation_poke.mp4` | the captured poke video (training camera 0) |
| `mat_prop_ckpt/carnation/` | the field recovered for the paper: `material_field.pt` (simulator format), `triplane_fields.pt` / `velocity_field.pt` (stage-2 / stage-1 state dicts), `selection.pt`, `sim_config.json` |

## foamedit bundle

| Path | Content |
|---|---|
| `powerfoam_ckpt/garden_v1/`, `powerfoam_ckpt/bonsai_foreground/` | the MipNeRF-360 garden (1.2 M primitives) and the bonsai captured on its own (300 k); `cameras_*.pt` include the image names the voting needs |
| `images/garden_v1/<image>.JPG` | the garden's 185 training photos at the checkpoint's resolution (1297×840), input of the mask generator |
| `masks/garden_v1/{vase,frond}/<image>.png` | binary 2D masks of the vase and its frond for all 185 garden views (Grounded-SAM, prompts "a vase with a plant." / "a dried palm frond.") |
| `selections/garden_v1/vase_and_frond_mask_paper.pt` | the selection used for the paper's composite |
| `selections/bonsai_foreground/tree_mask.pt` | the bonsai's tree without the pot (the part that is simulated), over the bonsai checkpoint's primitives |