language	library_name	license_name
en	nnll	MPL-2.0 + Commons Clause 1.0

negate
_{entrypoint synthetic image classifier}

A CLI tool and Python library for image processing and analysis of feature extraction to determine origin, whether synthetic/reconstructed or genuine.

Overview

We use a modern VAE to extract features from images generated using Diffusers, ComfyUI, Darkshapes tools (Zodiac/Divisor/singularity) and Google Nano-Banana. Using a modular pipeline that integrates a combination of feature extraction techniques such as spectral residual analysis and feature vectors extracted from the images, we train a Gradient Boosting Decision Tree model and its associated PCA transformer to distinguish between images of synthetic and human origin. Preliminary results demonstrate high accuracy in detecting synthetic images.

This repo provides a simple command‑line interface to invoke the tool and examples of integrating the library of predictions and metrics into other works. We make all attempts to follow continuous integration best practices for deploying and maintaining software, ensuring the code is readied for production environments.

Future work includes the development of an automated testing framework and evaluation suite, expanding the scope of research to include wider diversity of synthetic and original human-generated datasets, benchmarking against comparable methods, and exploring additional model architectures.

Requirements

• A dataset of images made by human artists with width and height dimensions larger than 512 pixels. This will serve as ground truth and should be placed in the /assets folder.
• A huggingface account that will be used to download models and synthetic datasets. Create an API Key at their website, then sign in with hf auth login.
• It is recommended to run negate on a GPU to ensure efficient processing and reduced training time.

Note

Our training results and visualizations were created with data provided consensually by generous artists at https://purelyhuman.xyz. We don't have and won't seek permission to share that dataset here.

Install

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

CLI:

Basic Syntax:

usage: negate [-h] {train,check} ...

Negate CLI

positional arguments:
  {train,check}
    train        Train model on the dataset in the provided path or `assets/`. The resulting model will be saved to disk.
    check        Check whether an image at the provided path is synthetic or original.

options:
  -h, --help     show this help message and exit

Training syntax:

usage: negate train [-h]
                    [-m {exdysa/dc-ae-f32c32-sana-1.1-diffusers,zai-org/GLM-Image,black-forest-labs/FLUX.2-dev,black-forest-labs/FLUX.2-klein-4B,Tongyi-MAI/Z-Image,Freepik/F-Lite-Texture,exdysa/mitsua-vae-SAFETENSORS}]
                    [path]

positional arguments:
  path                  Dataset path

options:
  -h, --help            show this help message and exit
  -m, --model {exdysa/dc-ae-f32c32-sana-1.1-diffusers,zai-org/GLM-Image,black-forest-labs/FLUX.2-dev,black-forest-labs/FLUX.2-klein-4B,Tongyi-MAI/Z-Image,Freepik/F-Lite-Texture,exdysa/mitsua-vae-SAFETENSORS}
                        Change the VAE model to use for training to a supported HuggingFace repo. Accuracy and memory use decrease from left to right

Check the origin of an image:

usage: negate check [-h] [-s | -g] path

positional arguments:
  path             Image or folder path

options:
  -h, --help       show this help message and exit
  -s, --synthetic  Mark image as synthetic (label = 1) for evaluation.
  -g, --genuine    Mark image as genuine (label = 0) for evaluation.

@article{darkshapes2026,
    author={darkshapes},
    title={negate},
    year={2026},
    howpublished={\url{https://github.com/darkshapes/negate}},
}