ViT ISR Tiny LSDIR
A Vision Transformer for ×4 image super-resolution — built entirely from scratch in PyTorch, trained on 76,716 real-world images from the LSDIR benchmark, a tiny model with less than million parameters
Fuente del modelo
- Tipo
- Other
- Tamaño del archivo
- 3,03–9,10 MB
- Versión
- vModel weights
- Actualizado
- 9 oct 2026
- Fuente
- Hugging Face
Descripción de la fuente
ViT-ISR-Tiny: Vision Transformer for ×4 Image Super-Resolution
PyTorch PSNR Params
A Vision Transformer for ×4 image super-resolution — built entirely from scratch in PyTorch, trained on 76,716 real-world images from the LSDIR benchmark, a tiny model with less than million parameters
What it does
Takes a low-resolution image and reconstructs a ×4 higher-resolution version using global self-attention — capturing long-range spatial relationships across the entire image rather than just local neighborhoods like CNNs do.
Architecture
The pipeline: patch embed → transformer blocks → reshape → PixelShuffle upsample → RGB output.
| Component | Details |
|---|---|
| Patch size | 2×2 → 1,024 tokens per image |
| Embedding dim | 64 |
| Transformer blocks | 6 (pre-norm, residual) |
| Attention heads | 4 |
| MLP hidden dim | 256 |
| Upsampling | 3× PixelShuffle ×2 (×8 total) |
| Parameters | ~786K · ~3 MB (fp32) |
Why pre-norm: LayerNorm before each sub-layer keeps the residual path clean, stabilizing training in deep stacks.
Why PixelShuffle: Learned upsampling — the model decides what to put in new pixels by redistributing channel information into space, rather than stretching with interpolation.
Why patch size 2: Finer tokens preserve more spatial detail for reconstruction. Larger patches are cheaper but lose the high-frequency information SR depends on.
Training
| Setting | Value |
|---|---|
| Dataset | LSDIR (76,716 train / 4,263 test) |
| Optimizer | AdamW (lr=2e-4) |
| Loss | L1 |
| Mixed precision | fp16 AMP |
| Batch size | 16 |
| Degradation | Bicubic ×4 downscaling |
| Hardware | RTX 4060 Laptop (8GB) |
Test PSNR: 23.30 dB — evaluated on the held-out test split (never seen during training or checkpoint selection).
Usage
from huggingface_hub import hf_hub_download
import torch
weights_path = hf_hub_download("Sathya77/ViT-ISR-Tiny-LSDIR", "sr_best.pt")
checkpoint = torch.load(weights_path, map_location="cpu")
model = ImageSRTransformer() # from model_architecture.py
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()
Files
| File | Description |
|---|---|
sr_best.pt | Best checkpoint (weights + optimizer state) |
model_weights.pt | Weights only — for inference |
config.json | Architecture config |
training_config.json | Training config and results |
Fuentes
1 fuente
Fuentes
1 fuenteVerificado 9 oct
Download files
Artefactos del modelo
2 artefactos
Artefactos del modelo
2 artefactosExtractos de fuentes
2 extractos
Extractos de fuentes
2 extractos--- tags: - image-super-resolution - vision-transformer - pytorch - computer-vision - super-resolution - lsdir datasets: - LSDIR metrics: - psnr --- # ViT-ISR-Tiny: Vision Transformer for ×4 Image Super-Resolution    A Vision Transformer for ×4 image super-resolution — built entirely from scratch in PyTorch, trained on 76,716 real-world images from the LSDIR benchmark, a tiny model with less than million parameters 👉 **[Try the Live Demo](https://huggingface.co/spaces/Sathya77/ViT-ISR-Tiny-LSDIR)** --- ## What it does Takes a low-resolution image and reconstructs a ×4 higher-resolution version using global self-attention — capturing long-range spatial relationships across the entire image rather than just local neighborhoods like CNNs do. --- ## Architecture The pipeline: patch embed → transformer blocks → reshape → PixelShuffle upsample → RGB output. | Component | Details | |-----------|---------| | Patch size | 2×2 → 1,024 tokens per image | | Embedding dim | 64 | | Transformer blocks | 6 (pre-norm, residual) | | Attention heads | 4 | | MLP hidden dim | 256 | | Upsampling | 3× PixelShuffle ×2 (×8 total) | | Parameters | ~786K · ~3 MB (fp32) | **Why pre-norm:** LayerNorm before each sub-layer keeps the residual path clean, stabilizing training in deep stacks. **Why PixelShuffle:** Learned upsampling — the model decides what to put in new pixels by redistributing channel information into space, rather than stretching with interpolation. **Why patch size 2:** Finer tokens preserve more spatial detail for reconstruction. Larger patches are cheaper but lose the high-frequency information SR depends on. --- ## Training | Setting | Value | |---------|-------| | Dataset | LSDIR (76,716 train / 4,263 test) | | Optimizer | AdamW (lr=2e-4) | | Loss | L1 | | Mixed precision | fp16 AMP | | Batch size | 16 | | Degradation | Bicubic ×4 downscaling | | Hardware | RTX 4060 Laptop (8GB) | **Test PSNR: 23.30 dB** — evaluated on the held-out test split (never seen during training or checkpoint selection). --- ## Usage ```python from huggingface_hub import hf_hub_download import torch weights_path = hf_hub_download("Sathya77/ViT-ISR-Tiny-LSDIR", "sr_best.pt")...
Source context: 0 downloads · 0 likes · Repo Sathya77/ViT-ISR-Tiny-LSDIR
