Downloads · 30 days
0
PolarFree-PSD-4ch-weights
nawta/PolarFree-PSD-4ch-weights
PolarFree-PSD-4ch-weights is a machine learning model from nawta. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Pre-trained weights for PolarFree model trained on PSD (Polarization Specular Dataset).
Access
Public
Updated Mar 4, 2026
Repo size
1.6 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pth1.6 GB · 100%
At a glance
- Access
- Public
- Created
- Jan 21, 2026
- Updated
- Mar 4, 2026
- SHA
- 4c2f96cc
Model card
From the Hugging Face model README
PolarFree PSD 4-channel Weights
Pre-trained weights for PolarFree model trained on PSD (Polarization Specular Dataset).
Important: 4-channel Polarization Format
PSD dataset originally has 12 polarization angles (30° intervals), but these weights were trained with 4-channel format (0°, 45°, 90°, 135°) to match standard polarization camera output.
Angle Conversion
PSD 12 angles → 4 angles:
- 0° → idx-01 (direct)
- 45° → interpolate(idx-02, idx-03)
- 90° → idx-04 (direct)
- 135° → interpolate(idx-05, idx-06)
Model Structure
stage1/
├── net_g_100000.pth # Generator (72MB)
└── net_le_100000.pth # Latent Encoder (2.5MB)
stage2/
├── net_g_50000.pth # Generator (72MB)
├── net_le_dm_50000.pth # Latent Encoder DM (2.4MB)
└── net_d_50000.pth # Denoising Network (10MB)
Performance
Test Set (50 samples)
| Stage | Iteration | PSNR (dB) | SSIM |
|---|---|---|---|
| Stage 1 | 100,000 | 34.85 | 0.9918 |
| Stage 2 | 50,000 | 34.45 | 0.9914 |
Inference Time (512x768)
| Stage | GPU (RTX 6000) | CPU |
|---|---|---|
| Stage 1 | 325 ms | ~31 sec |
| Stage 2 | 271 ms | ~22 sec |
Usage
from huggingface_hub import hf_hub_download
import torch
# Download Stage 1 weights
net_g_path = hf_hub_download(
repo_id="nawta/PolarFree-PSD-4ch-weights",
filename="stage1/net_g_100000.pth"
)
# Load
checkpoint = torch.load(net_g_path)
model.load_state_dict(checkpoint["params"])
License
Private repository - for research use only.
- Created
- Jan 21, 2026
- Updated
- Mar 4, 2026