🐍 Vision-Mamba: Lightweight SSM-based Vision Autoencoder

A lightweight Vision-Mamba network built using State Space Models (SSM) and a SimpleDecoder for image reconstruction and restoration tasks such as inpainting, colorization, and denoising.

The model leverages a Patch-based Encoder, MambaBlock (SSM) layers for long-range spatial reasoning, and a transposed-convolutional decoder for high-quality reconstruction.
Training integrates L1, SSIM, and VGG Perceptual losses for better structural and perceptual fidelity.

---

🚀 Features

• 🧩 Mamba-based Encoder Blocks – SSM-style modeling for efficient long-range spatial dependencies.
• 🎨 Simple Transposed-Conv Decoder – Reconstructs full-resolution RGB images.
• 📏 Multi-Loss Training – Combines L1, SSIM, and Perceptual losses for better color and texture recovery.
• ⚡ Lightweight & Fast – Designed for image colorization, denoising, and inpainting tasks.
• 🧠 PyTorch Implementation – Modular, readable, and easy to extend.

---

🖼️ Pipeline Overview

┌────────────┐     ┌────────────────────┐     ┌───────────────────────┐     ┌────────────────────────────┐     ┌───────────────┐
│ Input      │ --> │ Patch-based        │ --> │ MambaBlock (SSM)      │ --> │ Transposed-Conv Decoder    │ --> │ Output        │
│ Image      │     │ Encoder            │     │ Layers                │     │ (SimpleDecoder)            │     │ Image         │
└────────────┘     └────────────────────┘     └───────────────────────┘     └────────────────────────────┘     └───────────────┘

        ─────────────────────────────────────────────────────────────────────────────────────────
                              ↑ Losses applied during training (L1 + SSIM + VGG Perceptual)

---

🧩 License

This project is licensed under the MIT License — see the LICENSE file for details.