AI companies scraped billions of images from the open web without asking, and the SAG-AFTRA and WGA disputes made clear that creative jobs are already thinning because of it. If your portfolio sits on ArtStation, DeviantArt, a Squarespace site, or a Substack, it has almost certainly been touched by a crawler at some point.
The seven best apps for protecting art from AI training in this roundup will not stop every model from ever seeing your work, but they do raise the cost of copying it. We tested each on Windows, macOS, and Linux where supported, and cross-checked pricing, actively maintained builds, and what real illustrators, photographers, and voice actors on Bluesky and Reddit say holds up in 2026.
This list is for illustrators, photographers, writers, and voice actors who want to stop being training data for the next release.
What to look for in a desktop AI training protection app
Every tool on this list does one narrow thing well. The trick is matching each tool to how your work actually reaches the internet.
Five criteria are worth thinking about before you install anything.
How your work leaves the studio. JPEGs on a portfolio site need different defenses than raw voice takes uploaded to a client folder, or draft prose posted on a personal blog.
What kind of AI use worries you most. Style mimicry, image-to-image edits, dataset ingestion, and outright deepfakes each have a different mitigation, and no single app covers all four.
Hardware on hand. Cloaking a 4K piece with Glaze or Mist on a CPU-only laptop can take twenty minutes per image. A recent NVIDIA GPU or an Apple Silicon chip cuts that to well under a minute.
Timing. Some tools sit in front of your image so a scraper never sees it. Others alter the image so any scraper that does see it produces a broken training example.
Trust in the vendor. Glaze and Nightshade come from a university lab and publish their papers. Others sit behind a login. The trade-off matters if you post under a legal name.
Quick comparison
| App | Best for | Platforms | Free plan | Starting price | Source |
|---|---|---|---|---|---|
| Glaze | Cloaking visual style before posting | Windows, macOS, Linux | Full | Free | Academic (UChicago) |
| Nightshade | Poisoning scrapers that ignore opt-outs | Windows, macOS, Linux | Full | Free | Academic (UChicago) |
| Mist v2 | Open-source cloaking with GPU | Windows GUI, macOS and Linux CLI | Full | Free | GPLv3 |
| PhotoGuard | Immunising photos against AI edits | Windows, macOS, Linux (Python) | Full | Free | MIT |
| WebGlaze | Glazing without a capable GPU | Any desktop browser | Full (approved artists) | Free | Academic (UChicago) |
| Kudurru | Blocking scrapers at the hosting layer | Web dashboard, WordPress plugin | Individual tier | Free | Proprietary |
| Have I Been Trained | Auditing training sets and opting out | Any desktop browser | Unlimited | Free | Proprietary |
The apps
1. Glaze - Best overall for cloaking your visual style
The Glaze desktop app adds perturbations designed to disrupt style-mimicry training, so a diffusion model fine-tuned on the cloaked file learns a very different style than the one your eye sees. In the shipping build (Glaze 2.2, April 2026) a cloaked image ranges from visually identical at low intensity to a faint watercolor haze at high intensity, and every step runs locally so files never leave your machine. It is the default artist-facing option for style protection in 2026, with more than 8.5 million downloads across illustrators, comics artists, and concept designers since March 2023.
Where it falls short: the July 2025 LightShed paper from an academic team showed a targeted stripper can weaken Glaze protections when trained on many cloaked samples, and cloaking a 4K piece on a CPU-only laptop can crawl toward twenty minutes.
Pricing:
- Free: full functionality, all intensity levels, unlimited images.
- Paid: not offered.
Platforms: Windows 10 or 11, macOS 13 or newer on Apple Silicon or Intel, Linux via source build.
Download: Glaze
Bottom line: Anyone posting original visual art in 2026 should Glaze it before hitting publish, understanding that no cloak is permanent.
2. Nightshade - Best free companion to Glaze
Where Glaze protects an individual style, Nightshade goes on offense. It adds perturbations that push a scraper to associate the image with the wrong concept, so a model trained on many poisoned images returns broken generations for the poisoned prompt. The University of Chicago paper showed that fewer than a hundred poisoned samples can steer a single Stable Diffusion prompt off the rails, and downloads passed 2.5 million by mid-2026.
Where it falls short: Nightshade only helps if a scraper actually ingests your work. If your reach is small the poisoning effect on any given large model is negligible, and pairing Nightshade with Glaze roughly doubles the compute time per image.
Pricing:
- Free: full functionality.
- Paid: not offered.
Platforms: Windows 10 or 11, macOS 13 or newer, Linux via source build.
Download: Nightshade
Bottom line: Worth running if you post at volume, want to signal that your work is not passive training data, and can spare the compute time.
3. Mist v2 - Best open-source alternative to Glaze
Mist v2 is a Psyker Team project that publishes its full training code under GPLv3 and offers a Windows GUI plus a cross-platform Python CLI. It attacks the same image-encoder representations Glaze targets, adds perturbations designed to disrupt LoRA fine-tuning specifically, and can be batched from the command line for a full portfolio in one run. In head-to-head comparisons Mist v2 holds up against ControlNet-based mimicry attempts that pierce weaker cloaks.
Where it falls short: the perturbation is more visible than Glaze at equal strength, and the CLI expects a comfortable relationship with Python and a CUDA-capable GPU.
Pricing:
- Free: full source under GPLv3.
- Paid: not offered.
Platforms: Windows (GUI), macOS and Linux via Python CLI, CUDA GPU strongly recommended.
Download: Mist v2 on GitHub
Bottom line: Pick Mist v2 if you want a fully auditable open-source cloak and are comfortable at a terminal.
4. PhotoGuard - Best for photographers protecting a shoot from AI edits
MIT CSAIL’s PhotoGuard was built for a different threat model than Glaze: someone downloads your photo and runs it through Stable Diffusion inpainting to produce a deepfake or non-consensual edit. PhotoGuard immunises an image so any diffusion model that tries to edit it produces a warped, obviously broken result. The reference implementation is a Python notebook and a gradio demo you can run locally on a laptop with a supported GPU.
Where it falls short: PhotoGuard was published in 2023 and the underlying research has not shipped as a polished consumer app since. It also targets diffusion-model editing rather than initial training, so scrapers can still pull the raw file for other uses.
Pricing:
- Free: MIT-licensed source, no runtime charges.
- Paid: not offered.
Platforms: Windows, macOS, Linux (Python 3.9 or newer, CUDA GPU recommended).
Download: PhotoGuard on GitHub
Bottom line: Portrait photographers, wedding photographers, and photojournalists worried about downstream inpainting should immunise hero shots with PhotoGuard before publishing.
5. WebGlaze - Best for artists without a GPU
WebGlaze is the University of Chicago team’s browser version of Glaze, for illustrators on Chromebooks, older MacBooks, or Windows machines with integrated graphics. Processing happens on their servers instead of your CPU, which turns a twenty-minute local cloak into a coffee-break runtime. Access is invitation-based to keep bad actors out, and requests are approved by a small human team, usually within a few days.
Where it falls short: you have to apply and wait for approval, and the file leaves your machine for the duration of the cloak.
Pricing:
- Free: full functionality after approval.
- Paid: not offered.
Platforms: any modern desktop browser on Windows, macOS, Linux, or ChromeOS.
Download: WebGlaze
Bottom line: WebGlaze is the practical fallback when the desktop Glaze app is too slow for the hardware you have.
6. Kudurru - Best for blocking scrapers at the hosting layer
Kudurru is Spawning.ai’s active-defense network. It maintains a live list of IPs demonstrating scraping behaviour and lets rights holders refuse to serve those IPs or return decoy content instead. The desktop-side view is the WordPress plugin and the browser dashboard, so anyone hosting their own portfolio on WordPress, Ghost, or a self-hosted static site can put Kudurru in front of image requests without touching Glaze or Nightshade.
Where it falls short: Kudurru operates at the network layer, so it only helps for work hosted on domains you control. Portfolios on ArtStation or Instagram sit outside its reach.
Pricing:
- Free: individual accounts, WordPress plugin included.
- Paid: Studio tier for larger publishers, quoted by request.
Platforms: web dashboard on any desktop browser, WordPress plugin, self-hosted proxy for advanced setups.
Download: Kudurru
Bottom line: Set Kudurru up on any domain you own before you rely on image-level cloaks alone.
7. Have I Been Trained - Best for auditing what already leaked
Have I Been Trained is Spawning.ai’s public search over LAION-5B and the other open training sets behind Stable Diffusion. Drop in an image or a portfolio URL and it tells you whether your work is already inside the corpus. From the same page you can add pieces to Spawning’s Do Not Train registry, which signals opt-out to compliant scrapers and integrates with the ai.txt standard that platforms like DeviantArt and Shutterstock now honour.
Where it falls short: the registry only binds scrapers who choose to honour it, and LAION is one dataset among many. Anyone training off fresh common-crawl scrapes today may ignore the registry entirely.
Pricing:
- Free: unlimited image and URL searches, unlimited registry additions.
- Paid: not offered for individual creators.
Platforms: any modern desktop browser.
Download: Have I Been Trained
Bottom line: Audit yourself first, then decide which of the tools above matters most for the work that already leaked.
How to pick the right one
If you want the simplest option: install Glaze, run it on every new piece before you post, and stop worrying about the rest until you have time.
If you make a living from your visual style: layer Glaze and Nightshade on new work, and run Have I Been Trained against your back catalogue.
If you shoot people (portrait, wedding, journalism): PhotoGuard on hero shots is the single most important tool, since inpainting deepfakes are the sharpest immediate threat.
If you are on a Chromebook or older laptop: skip the local apps and use WebGlaze plus a Kudurru-guarded portfolio.
If you are a WordPress-hosting writer or photojournalist: Kudurru first, then Glaze or PhotoGuard on individual assets.
If you tried Glaze and hated the compute time: try Mist v2 for a similar defense with a slightly more visible cloak but much faster batches on a GPU.
FAQ
Can any tool truly stop AI from training on my art?
No, and any tool claiming otherwise is overselling. Glaze, Nightshade, Mist, and PhotoGuard add perturbations designed to disrupt model training or editing, but adversarial research (notably LightShed in July 2025) has shown some cloaks can be weakened by a targeted stripper. Combined defenses raise the cost of copying, they do not make it impossible.
Is Glaze safe to use on client work?
Yes for the finished deliverable you post publicly. Glaze does not modify your source file, and cloaked exports keep the visual look for a human viewer. Avoid cloaking proofs a client will edit further, because the perturbations can interact badly with heavy colour grading and lossy compression.
What is the best free app for protecting art from AI training?
Glaze from the University of Chicago is the default answer for illustrators. It is free, actively maintained, and the desktop app runs offline on Windows, macOS, and Linux. Nightshade is the strong second free tool that complements it, and Mist v2 is the best open-source alternative if you prefer inspectable code.
Do I need a GPU?
For Glaze, Nightshade, and Mist you get much better results with a recent NVIDIA GPU or an Apple Silicon chip. On a CPU-only laptop the local apps still work, but processing a single 4K image can take twenty minutes. WebGlaze is the workaround when the hardware is not there.
Does opting out on Have I Been Trained actually work?
For compliant model builders, yes. The Do Not Train registry is honoured by a growing list of platforms and by the ai.txt opt-out standard, but any scraper that ignores robots.txt today will ignore this too. Treat the registry as necessary but not sufficient.
What about voice actors and writers?
The perturbation tools above target images. Voice actors should watch the Spawning ecosystem, add explicit “no AI training” language to platform terms, and register clips on the Do Not Train registry. Writers can add ai.txt on domains they control and register texts through Spawning, but content-level perturbation for prose is still a research problem in 2026.