Spaces:
Running
Running
File size: 4,649 Bytes
f376a33 3262d11 f376a33 3262d11 f376a33 3262d11 f376a33 3262d11 f376a33 3262d11 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 | ---
title: SpectraGAN
emoji: 🖼️
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.31.0
app_file: app.py
pinned: false
license: apache-2.0
---
# 🖼️ SpectraGAN — Multi-Model Upscaler Comparison
A Gradio web app that lets you upscale an image with **multiple SR models simultaneously** and compare results side by side.
Supported models:
| Model | Architecture | Scale |
|-------|-------------|-------|
| Real-ESRGAN ×2 | GAN (residual-in-residual dense block) | ×2 |
| Real-ESRGAN ×4 | GAN (residual-in-residual dense block) | ×4 |
| SRCNN ×4 | Shallow 3-layer CNN | ×4 |
| HResNet ×4 | Deep residual network (EDSR-style) | ×4 |
| SR3 *(stub)* | Diffusion model | ×4 — see note below |
---
## 📑 Table of Contents
1. [Features](#features)
2. [Project Structure](#project-structure)
3. [Prerequisites](#prerequisites)
4. [Installation](#installation)
5. [Adding Your ONNX Models](#adding-your-onnx-models)
6. [Running Locally](#running-locally)
7. [SR3 Integration Guide](#sr3-integration-guide)
8. [Contributing](#contributing)
9. [License](#license)
---
## ✨ Features
- **Side-by-side comparison** — run up to 4 models at once, results displayed in a 4-panel grid.
- **Selective execution** — toggle any model on/off before running; unchecked models are skipped.
- **×8 post-resize** — optionally apply a bicubic ×2 pass on top of any ×4 result.
- **Tile-based inference** — large images are split into tiles matching each model's fixed input size, then stitched back together seamlessly.
- **Per-result download** — each panel has its own PNG download button.
- **Graceful degradation** — if a model file is missing (e.g. Drive ID not yet set), that panel is skipped without crashing the others.
---
## 📁 Project Structure
```
spectragan/
├── model/
│ ├── Real-ESRGAN_x2plus.onnx # auto-downloaded
│ ├── Real-ESRGAN-x4plus.onnx # auto-downloaded
│ ├── SRCNN_x4.onnx # you provide — see below
│ └── HResNet_x4.onnx # you provide — see below
├── app.py
├── requirements.txt
└── README.md
```
---
## ⚙️ Prerequisites
- Python 3.10+
- `git`
- A terminal / command prompt
---
## 🔧 Installation
```bash
git clone https://github.com/ParamAhuja/SpectraGAN.git
cd SpectraGAN
python -m venv .venv
source .venv/bin/activate # Linux/macOS
# .venv\Scripts\activate # Windows
pip install -r requirements.txt
```
---
## 🗂️ Adding Your ONNX Models
The Real-ESRGAN weights are downloaded automatically from Google Drive on first run.
For **SRCNN** and **HResNet** you need to:
1. Export your trained PyTorch model to ONNX:
```python
import torch
# SRCNN example
from srcnn import SRCNN
model = SRCNN()
model.load_state_dict(torch.load("srcnn.pth"))
model.eval()
dummy = torch.randn(1, 3, 128, 128)
torch.onnx.export(
model, dummy, "SRCNN_x4.onnx",
input_names=["input"], output_names=["output"],
dynamic_axes={"input": {2: "H", 3: "W"}, "output": {2: "H", 3: "W"}}
)
```
2. Upload the `.onnx` file to Google Drive and set **"Anyone with the link can view"**.
3. Copy the file ID from the share URL and update `DRIVE_IDS` in `app.py`:
```python
DRIVE_IDS = {
...
"srcnn_x4": "YOUR_SRCNN_DRIVE_FILE_ID_HERE",
"hresnet_x4": "YOUR_HRESNET_DRIVE_FILE_ID_HERE",
}
```
---
## 🚀 Running Locally
```bash
python app.py
```
Open `http://127.0.0.1:7860` in your browser.
---
## 🌀 SR3 Integration Guide
SR3 (Super-Resolution via Repeated Refinement) is a **diffusion model** — it cannot be exported to a static ONNX graph because its inference involves a variable-length denoising loop.
To add SR3:
1. Clone the reference implementation:
```bash
git clone https://github.com/Janspiry/Image-Super-Resolution-via-Iterative-Refinement
```
2. Place your trained checkpoint at `model/sr3_x4.pth`.
3. Add `torch` and `torchvision` to `requirements.txt`.
4. Write a wrapper in `app.py`:
```python
def run_sr3(input_img: Image.Image) -> Image.Image:
# load config + model, run the denoising loop, return result
...
```
5. Add `"sr3_x4"` to the `PANEL_KEYS` list and wire `run_sr3` into `compare_models`.
---
## 🤝 Contributing
Pull requests welcome. Please open an issue first to discuss significant changes.
---
## 📄 License
Apache 2.0 — see `LICENSE`.
---
## 👤 Author & Credits
- Real-ESRGAN by [xinntao](https://github.com/xinntao/Real-ESRGAN)
- SRCNN by Dong et al. (2014)
- HResNet / EDSR by Lim et al. (2017)
- SR3 by Ho et al. (2022) — [paper](https://arxiv.org/abs/2104.07636)
|