37 lines
2.7 KiB
Markdown
37 lines
2.7 KiB
Markdown
# Glaze - Style Protection Against AI Mimicry
|
||
|
||
**Glaze** is a tool developed by the University of Chicago that adds imperceptible perturbations to digital artwork. These perturbations protect an artist’s unique visual style from being mimicked by text-to-image generative models while remaining visually identical to human viewers. It is one of the most mature and widely adopted image-level defenses available to individual creators.
|
||
|
||
## Why Glaze Matters
|
||
|
||
AI image models are trained to extract and reproduce artistic styles. Once a style is learned, models can generate new works “in the style of” a specific artist without permission. Glaze disrupts this process at the feature level by shifting the image in CLIP-space so that the model associates the artist’s work with a different style. The change is invisible to humans but causes the model to fail at accurate style mimicry.
|
||
|
||
This technique directly counters the “style theft” problem highlighted in the primary dissertation and gives individual artists a practical, self-deployable tool that does not require cooperation from AI labs.
|
||
|
||
## How It Fits the Defense Stack
|
||
|
||
1. **Anubis + Nepenthes** - Web-level protection (prevent initial scraping).
|
||
2. **Canary tokens & active denial** - Attribution and real-time cost imposition.
|
||
3. **Glaze** (this document) - Post-scraping style protection for images that do get ingested.
|
||
4. **Nightshade** (`nightshade.md`) - Stronger concept-level poisoning (complementary to Glaze).
|
||
|
||
Glaze is the image-specific counterpart to the text and metadata techniques already documented. It is especially valuable for photographers, illustrators, and painters who publish work online.
|
||
|
||
## Key Benefits for Individuals
|
||
|
||
- **Invisible to humans** - No visible degradation of the original artwork.
|
||
- **Survives common preprocessing** - Robust against resizing, compression, and format conversion.
|
||
- **Easy to use** - Desktop application with simple “protect” workflow.
|
||
- **Free for non-commercial use** - Accessible to independent artists.
|
||
- **Open research** - Backed by peer-reviewed work (USENIX Security 2023).
|
||
|
||
## Official Resources
|
||
|
||
- Project: https://glaze.cs.uchicago.edu/
|
||
- Paper: “Glaze: Protecting Artists from Style Mimicry by Text-to-Image Models,” USENIX Security 2023
|
||
|
||
## Recommended Starting Point
|
||
|
||
Run Glaze on every publicly shared image before upload. Combine with daily canary tokens in metadata and the aggressive-bot conditional serving logic so that known scrapers receive Glazed + canaried versions. This creates a multi-layered defense that protects both the web presence and any images that are successfully stolen.
|
||
|
||
*Glaze is the primary style-protection tool in the image defense layer. It is designed to be used alongside Nightshade for stronger protection.* |