Cloak - AI Art Protection
free · quick · anonymous
Add invisible protection before you post your work.
Cloak layers subtle noise onto your image so that AI models detect harmful content (like watermarks) and refuse to train on your art. Created by Timothy via Claude Code.
Original click to upload
Drop or choose an image<br>Up to 30 MB
Image ready<br>No preview available. The image will still be protected.
Protected locked
Your result will appear here.<br>Protect an image to download it.
Settings
The default settings are recommended for the first run. Tweaking the settings are encouraged for subsequent runs.
Target Prompt
Short prompts steer model embeddings better.
Protection Strength<br>0.008
Higher values are stronger but form visible noise.
Processing Passes<br>250
More passes refine the protection but take longer.
Protect Image
Upload an image to begin.
Cosine Similarity Score
Cloak maximizes this score to steer AI models into classifying your image as the prompt. Scores above 0.4 are a strong match.
Original
Protected
Common Questions
with honest and transparent answers.
How does Cloak actually work? +<br>Cloak optimizes L-infinity perturbations using projected gradient ascent to maximize the cosine similarity score between a CLIP embedding and a target embedding. What does that mean? Big image-scraping pipelines use AI vision models to sort and filter what they collect, and they routinely discard images with watermarks. Cloak turns that habit against them: it reads your image the way a scraper's model would, then adds faint, carefully-shaped layers of noise that steer the model's reading toward the label "watermark" without changing how the image looks to humans. When a pipeline scores your image as watermarked, its usual response is to filter it out, so your work is less likely to end up in a training set. Pipelines are also likely to discard images with "blood" or "blurry" artifacts.
Sounds too good to be true. Is Cloak really free? +<br>Yes. Cloak is 100% free. No ads or data collection, either. This is a passion project. In return, there are a few limits in place. The maximum file size is 30 megabytes. The maximum image length is 10,000 pixels. The maximum processing passes is 300 steps. The maximum requests is 60 images per hour. These rules keep the service available to everyone.
Why should I use Cloak over other adversarial programs? +<br>Cloak fulfills a unique niche that other solutions do not. For example, Glaze assumes that your image will be trained upon, which is uncomfortable to artists. Meanwhile, Nightshade requires protected images to represent a big portion of the training dataset, which is not guaranteed. That being said, the best practice would be to use Cloak, Glaze, AND Nightshade to maximize protection.
What else can I do to protect my images from AI? +<br>Obtaining copyright allows you to legally control who uses your art. Adding important metadata, IPTC data, and C2PA credentials to your images also lowers its chances of being used to train AI. Some social media websites like Cara have special robots.txt policies that warn bot scrapers not to download images. Beyond Cloak, there are also other free anti-AI noise tools like Glaze, Nightshade, and Artshield.io available online. You can even apply actual watermarks to your art. Having more layers of protection increases the chances of keeping your art safe.
Are my images being collected?+<br>Your images are deleted after you receive the result. Frankly, a gallery of uploads would be tedious and expensive to maintain. Cloak also strips away important embedded metadata from your images, such as location and camera info.
What target prompts are the best?+<br>The best target prompts are based on images that AI training data tries to avoid. For example, recommended prompts include 'watermark', 'blood', 'gore', 'blurry', 'bad quality', 'explicit content', 'low-quality', 'passports', 'driver's license', 'low resolution', 'ai-generated', etc.
What can Cloak protect against? +<br>Cloak is designed against mass scrapers who use automated tools to process the countless images they collect. Applying protection raises the odds that your image is filtered out of large training datasets. Unlike normal watermarks, Cloak does not leave a visible stamp across your art at normal strength. Importantly, the protection endures even if the image undergoes cropping and resizing.
What can't Cloak protect against? +<br>Cloak cannot remove or protect images that have already been accepted into a training dataset. Cloak cannot guarantee that every model and pipeline will fall for the noise signal. Protection can weaken over time since models change over time. Cloak is also not a substitute for a true watermark, copyright, or legal protection.
Why is it taking so long? +<br>Cloak should only take a few seconds to a minute. The queue may be very busy, or your image may be too large. The page will update when the process is over, so feel free to keep it open in the...