README.md

metadata

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
2. Project Structure
3. Prerequisites
4. Installation
5. Adding Your ONNX Models
6. Running Locally
7. SR3 Integration Guide
8. Contributing
9. 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

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:

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:

DRIVE_IDS = {
    ...
    "srcnn_x4":   "YOUR_SRCNN_DRIVE_FILE_ID_HERE",
    "hresnet_x4": "YOUR_HRESNET_DRIVE_FILE_ID_HERE",
}

---

🚀 Running Locally

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:

   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:

   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
• SRCNN by Dong et al. (2014)
• HResNet / EDSR by Lim et al. (2017)
• SR3 by Ho et al. (2022) — paper