|
|
--- |
|
|
license: mit |
|
|
tags: |
|
|
- art |
|
|
- Adverse |
|
|
--- |
|
|
|
|
|
# ImageShield |
|
|
Digital Illustration Copyright Protection Tool for Watermarking and Unauthorized AI Training (Adversal Attack) |
|
|
|
|
|
For release please head for [Github](https://github.com/huzpsb/ImageShield/) |
|
|
|
|
|
[Preview Version] |
|
|
|
|
|
SCREENSHOT: |
|
|
 |
|
|
[Discord](discord.gg/FBfXQtMS7X) |
|
|
|
|
|
|
|
|
IMPORTANT NOTICES: |
|
|
1. This software is designed to protect digital illustrations from unauthorized AI training usage |
|
|
2. You must own the copyright to any images you process with this tool |
|
|
3. All processed images must be clearly marked as "Not for AI Training" |
|
|
|
|
|
USAGE REQUIREMENTS: |
|
|
- Working directory must not contain spaces or non-English characters |
|
|
- Input folder should contain only image files (no subfolders) |
|
|
- Output folder must exist and be empty before processing |
|
|
- Recommended image size: 500x500 to 2000x2000 pixels |
|
|
|
|
|
SYSTEM REQUIREMENTS: |
|
|
- CPU processing only (GPU version requires separate compilation) |
|
|
- Approximately 1GB memory usage per 1920x1080 image |
|
|
- Processing time: ~1 minute per 1920x1080 image on i7 processor (20x Faster than most tools alike!) |
|
|
|
|
|
FEATURES: |
|
|
- MeshTwist: Prevents AI feature extraction, primarily affects LoRA training |
|
|
- FGSM Protection: Disrupts AI image analysis and image-to-image generation |
|
|
- Watermarking: Adds difficult-to-remove watermarks |
|
|
- Gradient Injection: Interferes with neural network training |
|
|
|
|
|
NOTES: |
|
|
- Image transparency will be converted to 100% opacity during processing |
|
|
- Protection methods are not completely irreversible but increase difficulty for AI usage |
|
|
- Use the help icons (?) for detailed information about each setting |
|
|
|
|
|
© 2024 Huazhong University of Science and Technology |
|
|
Z.H.Gao @ AIA & JayceXu @ SSE |
|
|
|
|
|
For (partial)source please head for https://github.com/huzpsb/DeWm |
