Glaze vs Nightshade vs a Visible Watermark. Which Should Artists Use?
Glaze and Nightshade sit on a different risk than a visible watermark. They aim at bulk scraping for AI training after you post. Jingwei's visible watermark and attribution claim aim at someone who already has the file and uses AI to edit it or remove the mark. You can use them together. One does not replace the other.
Jingwei's own pages group Glaze and Nightshade. They do not publish a technical difference between those two tools. For that split, read Glaze's and Nightshade's own docs. What follows is only the line Jingwei actually draws, and what a visible watermark is for. Broader stacking lives in How to protect your art from AI. Style-learning detail is in Protect art from AI style mimicry.
What are Glaze and Nightshade for?
The FAQ says Glaze and Nightshade mainly target bulk scraping for AI training after a work is online. The homepage says they stop AI from training on your style. The same FAQ lists them as local apps that often need a strong GPU, take longer per image, and leave a more visible perturbation. Jingwei is the tool the FAQ points at when someone already holds the file.
The layered publishing workflow puts Glaze and Nightshade in a different category from watermarks, metadata, and verification. It does not score those two products. It does not claim Jingwei integrates with them or replaces them. A robots.txt rule or a site notice is also not a promise that hostile scraping will stop.
So picking between Glaze and Nightshade is not a question this site answers. Both sit on the training-scrape side of the line. Choose with their own documentation. Do not treat Jingwei as a substitute for that job.
What does a visible watermark do instead?
Corner logos, copyright bars, and small repeated marks are easy for an inpainter. They are regular, sit apart from the drawing, and leave small holes to fill. An AI eraser can wipe a typical semi-transparent overlay in about half a second. That failure is covered in Why ordinary watermarks fail against AI watermark removal.
Jingwei's visible layers are deterrents written into pixels. Displacement shifts pixels under the text, along edges and light-dark boundaries, so a removal model no longer has smooth neighbors to copy. A forced fill tends to leave smears, blocks, and warps. Push the redraw hard and the picture itself changes. Halftone, ASCII, emboss, blur bars, blur blocks, and tiled text are the other listed textures. Face emboss lock is for a face you do not want swapped or locally repainted. The aim is a dirtier, more expensive restore. Nothing on the page promises that restoration is impossible.
The watermark matrix shows protected files next to AI restoration attempts, with problem areas marked. The FAQ dates a set of those side-by-side tests to August 2026. Moiré and face-emboss samples are previews on Protect until they are re-exported. Use the matrix to see what a given layer looks like after a repair attempt. It does not publish a win rate.
A second layer sits beside the visible watermark. The JW declaration stores your name, whether the work is Original or AI-assisted, and usage limits you state. DWT frequency watermarks, LSB, EXIF/IPTC, and tracking anchors can sit with it. Verify reads those Jingwei signals and can summarize C2PA if the file already has it. Protect does not write C2PA. A hit means the signal is still readable. It is not legal proof of ownership.
Invisible marks do not raise the cost of a clean look. Someone can still walk away with a picture that looks unprotected. If they never redraw it, you may still read the claim later. Heavy compression, repeated saves, and a full repaint can weaken or remove it. Screenshots sometimes still verify, and sometimes do not. How to place those layers day to day is in How to watermark your art.
When should you use which?
Match the tool to the risk you actually have. The FAQ comparison is the one to trust.
Glaze or Nightshade, when the risk is bulk scraping for training after you post. Plan on a local app, often a strong GPU, a slower export, and a more visible perturbation. Jingwei does not run those tools for you.
Jingwei, when the risk is a person who already has the file. That covers an AI edit, a fusion, a local redraw, or a watermark wipe. It runs in the browser, uses classical image processing, ships no deep-learning model files, takes about 30 seconds, and deletes the upload after processing. No account.
Both, when you have both risks. The FAQ says you can use all three. Do that as three separate steps, each for its own job. This site does not say that running Protect will preserve a Glaze or Nightshade perturbation, or the reverse. It also does not say the combination is pointless. Keep a copy from each step until you have checked the result yourself.
Inside Jingwei, pick the strength on purpose. Quick Credit is the daily default. You get your name, a JW claim, and a faint character field or a light displacement signature (about 1% of pixels on the site's textured example). That setting is meant to stay verifiable without looking ugly. It is not the strongest wipe resistance. Manual does not share those sliders. Use it to tile displacement, add emboss, add a blur bar, or lock a face, then compare with the matrix. Download PNG. JPEG weakens invisible layers.
A posting order that keeps the two jobs separate
1. Keep the unmarked master offline.
2. If training scrapes are the risk you care about, run the anti-training tool you chose, using that project's own settings.
3. Export the public file you are willing to show. Do not upload the largest master.
4. If edits, reposts, and watermark removal are the risk, run that export through Protect.
5. Verify the PNG once. Save the URL, the time, and a screenshot when you post.
Step 2 and step 4 are different exports. Nothing on the site merges them into one button.
What none of these tools can promise
None of them can promise that every future tool will fail. Keep the originals.
A visible layer does not block downloads or screenshots. It raises the cost of erasing the mark and tends to leave more artifacts after a forced cleanup. An invisible claim does not survive every compress, transcode, or screenshot, and it is not standalone proof of ownership. Glaze and Nightshade, as Jingwei describes them, do not solve the case where someone already has your file and is editing that file.
C2PA does not fill the gap either. If a camera or editor already wrote Content Credentials, Verify may summarize them. Jingwei's protect flow does not generate them, and a credential is not an independent verdict that a human made the picture or that you own it.
If you only post occasionally and you mostly fear a repost with the signature wiped, start with Protect in Manual and a visible layer you have looked at on the matrix. If you mostly fear your style being scraped for training, start with Glaze or Nightshade and treat a watermark as a second, separate step. If you fear both, do both, and do not expect either one to cover the other's failure.
Does Jingwei replace Glaze or Nightshade?
No. Glaze and Nightshade target bulk scraping for training. Jingwei targets someone who already has the file and runs an AI edit, fusion, or local redraw. The site says you can use all three. It does not say Jingwei integrates with or replaces the other two.
Should I run Glaze and Jingwei on the same file, and in what order?
Use each tool for its own risk. This site does not say whether Protect preserves a Glaze or Nightshade perturbation, or whether those tools preserve Jingwei's marks. Keep a copy from each step and check it. Do not assume one pass does both jobs.
If I only add an invisible watermark, am I protected from AI edits?
No. Invisible layers are for checking attribution later, and they can be weakened by compression or a full repaint. Visible layers are what raise the cost of a clean restore. A Verify hit is a technical clue, not legal proof of ownership.
Protect the next public file
If training scrapes are the fear, keep Glaze or Nightshade in the stack. If the posted file is the fear, run that export through Jingwei before it goes live. When both matter, do both as separate steps.
https://jwprotect.com/protect