Downloads · 30 days
0
dlp-nppe3-rrdb-sr
zenfacehug/dlp-nppe3-rrdb-sr
dlp-nppe3-rrdb-sr is a machine learning model from zenfacehug. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
RRDB-based (ESRGAN backbone, no GAN loss) model for joint low-light denoising and 4x super-resolution, trained for the DLP 26T2 NPPE3 Kaggle competition.
Access
Public
Updated Sep 8, 2026
Repo size
30.4 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pth30.4 MB · 100%
At a glance
- Access
- Public
- Created
- Sep 6, 2026
- Updated
- Sep 8, 2026
- SHA
- 3ea73fce
Model card
From the Hugging Face model README
RRDBNet — 4x Super-Resolution + Denoising (DLP NPPE3)
RRDB-based (ESRGAN backbone, no GAN loss) model for joint low-light denoising and 4x super-resolution, trained for the DLP 26T2 NPPE3 Kaggle competition.
Architecture
- RRDBNet: nf=64, nb=10, gc=32, 4x pixel-shuffle upsampling (two 2x stages)
- Trained on 256x256 HR / 64x64 LR patches with random flip + 90-degree rotation augmentation
Training
- Loss: L1, then a short L2 (MSE) fine-tune with EMA weight averaging to directly target PSNR
- Optimizer: AdamW, cosine LR schedule
- Validation metric tracked in two ways: full-RGB PSNR, and the competition's actual grading transform (grayscale, every-8th-flattened-pixel)
Results
- Local val PSNR (full RGB): ~38.6 dB
- Public leaderboard PSNR: {public_score_placeholder} <!-- fill in your actual number -->
Usage
from model_def import load_model
model = load_model("model.pth")
- Created
- Sep 6, 2026
- Updated
- Sep 8, 2026