negate

entrypoint synthetic image classifier

About

A command-line tool and Python library for processing and analyzing images, extracting Laplacian residuals to measure fractal and texture complexity, and other comparative analysis methods to discriminate synthetic images from real ones.

Test Results

Demonstration of the provided test results and visualizations on our synthetic darkshapes/a_slice dataset and private works of human origin provided by consent from the generous artists at https://purelyhuman.xyz.

Install

Important

Requires uv

git clone https://github.com/darkshapes/negate.git
cd negate
uv sync

_macos/linux

source .venv/bin/activate

_windows

Set-ExecutionPolicy Bypass -Scope Process -Force; .venv\Scripts\Activate.ps1

Test:

Add human-origin assets to assets/real

python -m tests.test_extract -v

Scan A Folder or File

usage: negate [-h] [-i INPUT] [-g] [-v] [-o OUTPUT]

Extract Laplacian residuals from images.

options:
  -h, --help           show this help message and exit
  -i, --input INPUT    Input folder or individual image.
  -g, --graph          Graph the distribution of residuals on a plot
  -v, --verbose        Enable verbose output.
  -o, --output OUTPUT  (OPTIONAL) Output folder for residuals.

Train Model:

python -m negate.train

Call from another application

import asyncio
from pathlib import Path

from negate import ResidualExtractor, flag_origin

image_path = "some/image/path.jpg"
output_folder = ".output"

residual_extractor = ResidualExtractor(image_path, output_folder, verbose=False)

async def async_main() -> tuple[list,list]:
    fractal, texture = await residual_extractor.process_residuals()
    return residual_extractor

residual_extractor = asyncio.run(async_main())
flag_origin(residual_extractor)