LoRA Strength Analyzer

A Python script to analyze images generated using a LoRA (Low-Rank Adaptation) model applied at various strength levels. This tool helps determine an optimal strength for a given LoRA by evaluating image quality and similarity to control images.

(Note: Gradio webUI coming soon)

🧠 1. Structural Similarity Index (SSIM)

SSIM measures the similarity between two images in terms of:

• Luminance
• Contrast
• Structural features

It’s computed as:

$$ \text{SSIM}(x,y) = \frac{(2\mu_x\mu_y + C_1)(2\sigma_{xy} + C_2)}{(\mu_x^2 + \mu_y^2 + C_1)(\sigma_x^2 + \sigma_y^2 + C_2)} $$

Where:

• $\mu_x, \mu_y$: mean pixel intensities
• $\sigma_x, \sigma_y$: standard deviations
• $\sigma_{xy}$: cross-covariance
• $C_1, C_2$: constants to stabilize division

Interpretation:

• SSIM ≈ 1.0 → Very similar (minimal LoRA effect)
• SSIM ≪ 1.0 → Significant difference (strong LoRA effect)

🔍 2. BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator)

BRISQUE estimates the perceptual quality of an image without needing a reference. It uses machine learning and natural scene statistics to assess artifacts and distortions.

How it works:

• Extracts statistical features from image patches
• Feeds them into a pretrained model (typically SVM)
• Outputs a quality score

Interpretation:

• Lower score → Better image quality
• Higher score → More visible artifacts, noise, or degradation

🎯 Goal

By analyzing both:

• SSIM (similarity to original)
• BRISQUE (perceptual quality)

The tool helps you:

• Detect the best LoRA strength for subtle or strong stylistic changes
• Avoid over-strengthening that introduces artifacts
• Maintain good image quality while applying desired effects

Features

• Calculates BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator) score for each LoRA-strength image. Lower BRISQUE scores generally indicate better perceptual quality.
• Calculates SSIM (Structural Similarity Index Measure) between each LoRA-strength image and its corresponding control image (or a single control image). An SSIM score of 1.0 means identical.
• Provides a summary table of scores for all tested strengths.
• Suggests a "best" strength based on the lowest BRISQUE score (preferring lower strength in case of a tie).
• Supports using either individual control images for each strength or a single control image for all strengths.

Prerequisites

• Python 3.7+
• Git (for cloning)

Setup

1. Clone the repository:

   git clone https://github.com/Raxephion/loRA-Strength-Analyser.git
   cd lora-strength-analyzer

2. Create and activate a virtual environment:

   • Using venv:

     python -m venv venv
     # On Windows
     .\venv\Scripts\activate
     # On macOS/Linux
     source venv/bin/activate

   • Using conda:

     conda create -n lora_analyzer python=3.9 # Or your preferred Python 3.x version
     conda activate lora_analyzer

3. Install dependencies:

   pip install -r requirements.txt

4. Configure Script: Open lora_strength_analyzer.py in a text editor.

   • Update Author Information (Optional but good practice):

     # Modified to check LoRA Strength on [Your Modification Date] <--- UPDATE THIS
     # @author: raxephion (Original), [Your Name/Handle for modification] <--- UPDATE THIS

   • MUST Update Paths: Adjust the following placeholder paths to your actual directory locations:

     LORA_STRENGTH_IMAGES_DIR = Path(r"C:\...\LoRA_Training\lora_strength_images") # UPDATE THIS
     CONTROL_IMAGES_DIR = Path(r"C:\...\LoRA_Training\control_images") # UPDATE THIS

   • Adjust LoRA Strength Parameters: Configure the range and step for LoRA strengths:

     MIN_LORA_STRENGTH = 0.1
     MAX_LORA_STRENGTH = 1.0
     LORA_STRENGTH_INCREMENT = 0.1
     STRENGTH_DECIMAL_PLACES = 1 # For formatting e.g., 0.1, 1.0

   • Verify Image Naming Convention: The script assumes image names like:

     • LoRA strength images: strength_0.1.png, strength_0.2.png, ... Modify LORA_IMAGE_PREFIX and IMAGE_EXTENSION if yours differ.

   • Control Image Configuration: You have two options for control images:

     1. Multiple Control Images: (Default) One control image per LoRA strength tested.
        • Control images named: control_01.png, control_02.png, ... where control_01.png corresponds to MIN_LORA_STRENGTH, control_02.png to the next strength, and so on.
        • Modify CONTROL_IMAGE_PREFIX if needed.
        • Keep USE_SINGLE_CONTROL_IMAGE = False.
     2. Single Control Image: Use the same control image for all LoRA strengths.
        • Set USE_SINGLE_CONTROL_IMAGE = True.
        • Specify SINGLE_CONTROL_IMAGE_NAME = "your_control_image.png" (update this name).
        • Ensure this image exists in your CONTROL_IMAGES_DIR.

Usage

Once configured, run the script from your terminal:

python lora_strength_analyzer.py