liantian's picture
docs: clarify WorldEdge fixed model and preserve historical checkpoints
57e4213 verified
|
Raw History Blame Contribute Delete
4.44 kB
---
license: other
library_name: pytorch
tags:
- worldedge
- 3d
- articulation
- part-segmentation
base_model: rayli/Particulate
---
# WorldEdge — articulation module
**Use [`model.pt`](./model.pt). This is the single recommended checkpoint for all input meshes.**
WorldEdge adapts [PARTICULATE](https://github.com/RuiningLi/particulate) for its articulated-asset module. WorldEdge-Art names our constructed asset dataset; it is not a separate checkpoint to choose at inference. The articulation network retains the official 16-part B configuration and uses frozen PartField features.
The recommended file is a clean alias of the previously delivered `R7-WiSE050-lw/model.pt`, with **identical bytes**. The old `lw` suffix is historical: **do not select a different checkpoint for furniture, appliances, self-constructed assets, or unknown sources.** The same fixed checkpoint was evaluated across all 456 objects in the current released-versus-WorldEdge comparison.
| Current model | Value |
|---|---|
| Download | [`model.pt`](./model.pt) |
| Size | 605,384,630 bytes (605.4 MB) |
| SHA-256 | `2747c17bd0c8dedc07b7e45de92d802ce998cc4b6098f7b9532ab92d77bad2df` |
| Configuration | [`particulate-B.yaml`](./particulate-B.yaml), at most 16 parts |
| Inference source selection | None; one source embedding is already folded into the parameters |
| Machine-readable index | [`MODEL_INDEX.json`](./MODEL_INDEX.json) |
## Download
```bash
hf download liantian/particulate-hinge-final model.pt particulate-B.yaml MODEL_INDEX.json \
--local-dir models/worldedge-articulation
```
Verify the downloaded checkpoint:
```python
import hashlib
from pathlib import Path
path = Path("models/worldedge-articulation/model.pt")
assert hashlib.file_digest(path.open("rb"), "sha256").hexdigest() == (
"2747c17bd0c8dedc07b7e45de92d802ce998cc4b6098f7b9532ab92d77bad2df"
)
```
PartField is a **separate required feature-encoder dependency**, not another WorldEdge candidate. In a configured PARTICULATE/WorldEdge checkout:
```bash
hf download mikaelaangel/partfield-ckpt model_objaverse.ckpt \
--local-dir particulate/PartField/model
```
## Use in WorldEdge
From the project checkout, use its integration entry point with the downloaded `model.pt`:
```bash
python tools/infer_delivered.py \
--input-mesh /absolute/path/to/object.glb \
--checkpoint /absolute/path/to/models/worldedge-articulation/model.pt \
--up-dir=Y \
--output-dir /absolute/path/to/new-object-output
```
`--up-dir` describes the actual input coordinate system: glTF/GLB commonly uses Y-up; our processed OBJ assets use Z-up. The output directory must be new. CUDA and the project dependencies are required. See the project manual for setup, export, and per-object checks. The folded checkpoint is also compatible with the official PARTICULATE `infer.py` and its 16-part B configuration.
## Historical files
These remain available for provenance and earlier experiments. They are **not alternatives to select by input dataset**.
| Existing path | Status |
|---|---|
| `R7-WiSE050-lw/model.pt` | Legacy path for the **same current recommended weights** |
| `R7-WiSE050-pn/model.pt` | Historical source-conditioned export; not the fixed model used across all current datasets |
| `R7-WiSE050-own/model.pt` | Historical source-conditioned export; not the fixed model used across all current datasets |
| `F/model.pt` | Earlier mixed-source continued-training model |
| `F-WiSE050/model.pt` | Earlier 50:50 released/F parameter average |
[`MODELS.json`](./MODELS.json) retains historical training and export provenance; [`MODEL_INDEX.json`](./MODEL_INDEX.json) identifies the current recommendation. Historical server paths and generation manifests do not imply additional downloadable checkpoints.
## Evaluation scope
The current fixed model improves segmentation on the evaluated Lightwheel and WorldEdge-Art subsets. Corrected motion accuracy declines on PartNet-Mobility; it is not uniformly better on every dataset or metric. The cohorts informed development, and the measurements do not establish collision-free motion or physical task success. The model weights are unchanged by this repository organization.
## Terms
License remains `other`. Training data include PartNet-Mobility (non-commercial) and GRScenes (CC BY-NC-SA). The existing restriction on redistribution remains in place; this repository organization does not grant additional rights.