Hy-MambaIR

A hybrid Mamba-Transformer framework for cross-scale digital rock reconstruction in strongly heterogeneous conglomerates.

Overview

Hy-MambaIR is a hybrid Mamba-Transformer framework for cross-scale digital rock reconstruction in strongly heterogeneous conglomerates. The method is designed to improve both image-level reconstruction fidelity and physically meaningful structural consistency for engineering-oriented characterization.

The framework follows a local-to-global serial design:

• a local Transformer branch captures fine pore textures and boundary details
• an attentive state-space branch models long-range structural dependencies
• a cascaded structural refinement block (SCAB) enhances structural fidelity on global feature representations

Repository Contents

This repository includes:

• the Hy-MambaIR implementation
• training and evaluation scripts
• configuration files for training and evaluation
• one released checkpoint
• representative demo LR/HR patch pairs
• environment and dependency specifications

This repository does not include raw full micro-CT volumes or restricted source datasets.

Method Components

The current implementation uses the following module terminology:

• H-ASSB (HASSB in code): hierarchical local-to-global hybrid block
• H-SSM (HSSM in code): attentive state-space global modeling branch
• SCAB (SCAB in code): spatial-channel attention refinement block
• DyT (DyT in code): optional local-branch normalization variant

Environment

Validated local environment:

• Python 3.10.19
• PyTorch 2.10.0+cu128
• CUDA 12.8

Key dependencies:

• basicsr
• mamba_ssm
• causal_conv1d
• lpips
• timm

You can recreate the environment using either:

• environment.yml
• requirements.txt

Repository Structure

Hy-MambaIR/
├── README.md
├── LICENSE
├── environment.yml
├── requirements.txt
├── .gitignore
├── .gitattributes
├── configs/
├── scripts/
├── core/
├── checkpoints/
└── demo_data/

Quick Start

1. Install dependencies

Using Conda:

conda env create -f environment.yml
conda activate hymambair

Using pip:

pip install -r requirements.txt

If you cloned the repository with the released checkpoint, make sure Git LFS assets are available:

git lfs install
git lfs pull

2. Run a minimal smoke test

python scripts/smoke_test.py

3. Evaluate the released checkpoint on the demo subset

python scripts/evaluate.py --config configs/eval_demo.yaml --weights checkpoints/Hy-MambaIR_x4_main.pth --output results/eval_result.json

Checkpoint

The repository currently provides one released checkpoint:

• checkpoints/Hy-MambaIR_x4_main.pth

This checkpoint supports evaluation and smoke testing of the released x4 setup under the provided pipeline.

Demo Data

The repository includes four representative LR/HR demo patch pairs under:

• demo_data/LR/
• demo_data/HR/

These files are provided as a minimal reproducibility subset and do not replace the full dataset.

Data Availability

Raw micro-CT data are not included in this repository. Access may be subject to institutional approval and applicable data-use restrictions.

Citation

Citation information will be provided after peer review.

Contact

For technical questions regarding code usage, checkpoint issues, or dataset access, please contact the project maintainers.