PLSR / README.md
yxliu's picture
First commit
c4d09bb verified
|
Raw History Blame Contribute Delete
1.65 kB
---
base_model: microsoft/TRELLIS.2-4B
base_model_relation: adapter
tags:
- 3d
- super-resolution
- plsr
---
# PLSR Models
PLSR progressively refines the geometry and PBR materials of coarse 3D objects
using reference images. This repository contains two models built on TRELLIS.2:
| Model | Purpose | Weights |
| --- | --- | --- |
| Shape | Geometry super-resolution | `shape/shape_adapter.safetensors` |
| PBR | Material super-resolution | `pbr/pbr_adapter.safetensors` |
Each weight file contains the trained LoRA and PLSR extension parameters.
The corresponding `shape/config.json` and `pbr/config.json` define the model
architecture and inference parameters. `pipeline.json` specifies the component
paths and their TRELLIS.2 base weights.
The PLSR pipeline loads the TRELLIS.2 base weights, then applies these adapters.
TRELLIS.2 and DINOv3 are loaded from their own repositories.
## Load from local directories
Install the PLSR inference code and its dependencies, then download this model
repository and the TRELLIS.2 and DINOv3 repositories. Use their local directories:
```python
from plsr import PLSRPipeline
pipeline = PLSRPipeline.from_pretrained(
"/path/to/PLSR_release",
trellis2_repo_id="/path/to/TRELLIS.2-4B",
dino_repo_id="/path/to/dinov3-vitl16-pretrain-lvd1689m",
)
```
Each path points to the root of a downloaded Hugging Face repository.
The TRELLIS.2 and DINOv3 arguments also accept Hugging Face repo IDs:
`microsoft/TRELLIS.2-4B` and `facebook/dinov3-vitl16-pretrain-lvd1689m`.
For installation instructions and inference examples, see the
[PLSR code repository](https://github.com/LIU-Yuxin/PLSR).