obaskly/NeuralBleach
obaskly/NeuralBleach has 26 GitHub stars and 1 forks. GitHot measured 0 stars in 24 hours, 0 in 7 days, and 0 in 15 days. Last checked 2026-11-19 06:36 UTC. Last push 2026-02-04.
측정된 스타 증가
- in 24h
- 0
- in 7 days
- 0
- in 15 days
- 0
- in 1 month
- 업데이트 지연
서로 겹치는 기간의 순증가량이며 일별 합계가 아닙니다. 측정값이 없으면 비워 둡니다. 스타는 관심도를 나타내며 품질을 보장하지 않습니다.
소개
A specialized tool for removing invisible AI watermarks, such as Google SynthID, from images.
창작 및 미디어README.md
Original🧪 NeuralBleach
"If the pixels can dream, they can be made to forget."
NeuralBleach is a proof-of-concept utility for bypassing frequency-domain watermarking. NeuralBleach leverages a generative diffusion pipeline to overwrite hidden signatures by regenerating high-frequency noise. Unlike standard denoising tools, it employs edge-detection (ControlNet) to maintain semantic integrity during the reconstruction process.
Features
1. The "Launder" Protocol (Signal Disruption)
Uses Realistic Vision V5.1 (a photorealistic finetune of Stable Diffusion) to "re-dream" the image pixels.
- Why: Robust watermarks rely on specific pixel-grid coherence. By regenerating the image with a
strengthof 0.20 - 0.25, we force a new noise distribution that overwrites the original hidden signal without the "potato face" distortion of older models.
2. ControlNet "Structure Lock"
Standard AI tools destroy text and small details. NeuralBleach uses ControlNet (Canny Edge Detection) to create a "wireframe" of your image before processing.
- Result: The AI is forced to paint inside the lines. Text remains readable, logos stay sharp, and faces maintain their identity, even while the invisible pixel data is scrubbed.
3. The "Analog" Humanizer (Visual Spoofing)
Detectors don't just look for code; they look for "digital perfection." This module injects physical camera imperfections:
- Chromatic Aberration: Micro-shifts Red/Blue channels at the edges to simulate lens light refraction.
- Film Grain Simulation: Adds ISO-400 equivalent luminance noise to break the "AI smoothness."
4. Metadata Nuke
Automatically strips all EXIF, XMP, and C2PA (Content Credentials) headers, ensuring no digital signature remains in the file container.
Quick Start (Google Colab)
Don't have a GPU? Run Neural-Bleach in the cloud for free.
- Open Google Colab.
- Change Runtime to T4 GPU.
- Copy/Paste the script from
neural_bleach.colab. - Run and upload your image.
🛠️ Local Installation
If you have a GPU (NVIDIA RTX 3060 or better recommended), you can run this locally.
Prerequisites
pip install -r requirements.txt
Usage
# Basic Usage (Default Strength 0.20)
python neural_bleach.py image.png
# Aggressive Mode (If watermark persists)
python neural_bleach.py image.png --strength 0.25 --out output_image.jpg
Technical Pipeline
| Stage | Process | Technical Goal |
|---|---|---|
| 0. Pre-Processing | LANCZOS Upscale (1.2x) |
Increase pixel density for better facial reconstruction. |
| 1. Wireframing | OpenCV Canny |
Create a structural map to lock text & edges. |
| 2. Laundering | Realistic Vision + ControlNet |
Overwrite frequency-domain watermarks while respecting the wireframe. |
| 3. Humanization | np.roll + Gaussian Noise |
Defeat visual "smoothness" classifiers. |
| 4. Sterilization | Image.save(quality=95) |
Strip C2PA/Metadata & Header signatures. |
⚠️ Disclaimer
Educational Purposes Only. This tool is intended for cybersecurity research, CTF (Capture The Flag) competitions, and studying the robustness of digital watermarking technologies. The author is not responsible for misuse of this tool to bypass copyright protection or violate Terms of Service of generative AI platforms.
Example
| NanoBanana Pro | NeuralBleach |
|---|---|
POC
https://github.com/user-attachments/assets/639432a4-56c5-4f18-91bf-30915d56817e
License
Distributed under the MIT License.