Start here
What the Stable Diffusion watermark remover does
Drop an image into the panel above and it goes into our media workflow, where the check and the edit run together. This page exists to tell you something first, because the phrase Stable Diffusion watermark remover covers two problems that have nothing to do with each other and only one of them is a picture problem.
The first is a visible logo sitting on the image, added by whichever site you generated on. The second is an invisible payload written into the pixel data by the reference Stable Diffusion code itself. They look nothing alike, they are removed in nothing like the same way, and the commercial tool page we read most closely for this term never mentions the second one at all.
The distinction
The two Stable Diffusion watermarks, in plain terms
Get this right before you do anything, because it decides whether you have a job at all.
The invisible one is documented, not a rumour. The library used is invisible-watermark, and in the CompVis reference code the encoder is set up and applied by a put_watermark call inside txt2image.py. The payload it writes is the literal string StableDiffusionV1.
- A visible logo
- A brand mark in a corner of the frame. Nothing to do with the model. It comes from the hosted site you generated on, and it is the thing our workflow can edit.
- An invisible payload
- Bits written into the image's frequency coefficients by the invisible-watermark library. You cannot see it, it does not affect how the image looks, and no brush tool will touch it.
- A model signature
- A style artefact, not a watermark. Some custom checkpoints produce a recurring corner flourish because their training data was full of signed artwork. That is an inpainting job like any other.
How to
How to remove a Stable Diffusion watermark in three steps
The edit runs in our media workflow, straight from the panel above.
- Step 1
- Look at the image. If you can see a logo, it is a visible mark from the platform you used. If you cannot see anything, nothing on this page will make the image look different.
- Step 2
- Upload the file. Detection and the edit run together, one image at a time, and the result comes back as a new rendering rather than the original with a patch on it.
- Step 3
- Compare the result against the original at full size, paying attention to whatever detail sat underneath the mark. That is where a regeneration is most likely to differ.
Worth knowing
What survives editing and what does not
The invisible-watermark project publishes its own robustness results, and they are not what most people assume. In its own words the methods are not robust to resize or aspect-ratio-changed crop, but are robust to noise, colour filter, brightness and JPEG compression.
Read that again, because it inverts the usual instinct. Saving the file again as a JPEG keeps the mark. Running it through a colour filter keeps the mark. Its published test table shows the frequency-based methods failing at a 50% resize and failing under rotation. It also records a known defect: none of the algorithms performs well on web page screenshots or on posters with a homogeneous background colour.
The case where you do nothing
If you run Stable Diffusion yourself, there may be nothing to remove
Two of the repositories that turn up on this search exist only to switch the watermark step off at source, and the CompVis issue thread shows how small the change is: comment out the encoder setup and the put_watermark calls, and the reference script stops marking anything. A local AUTOMATIC1111 or ComfyUI install may already be producing unmarked files, and no hosted platform is adding a logo to an image you generated on your own machine.
One honest caveat. The person who opened that issue asked whether anything else in the codebase also marks the output, and the thread contains no definitive answer. We are not going to invent one.
Before you remove anything
Removing a watermark is not permission to use the image
This is the part the competing tool pages leave out entirely. Taking a mark off an image does not transfer any right to use it, and modifying an image is prohibited outright by plenty of licences. A watermark is often the licence being enforced, and removing it changes the picture, not the terms.
The use this page is written for is your own generated output: images you produced from your own prompt on a platform whose marks you are entitled to take off. Clearing someone else's branding from their work is a different activity and is not what we are helping with.
The straight answer
Does this guarantee a clean Stable Diffusion image?
No. Our workflow regenerates the image, which means it re-renders the file rather than lifting a sticker off the original. On a mark sitting over flat sky or plain background you will usually not be able to tell. Over fine detail, text, or a face, a regeneration can differ from the original in ways you will notice.
And a regeneration is not a promise about any invisible payload. We do not tell you what a decoder will report afterwards, ours or anyone else's, because nobody can honestly make that promise about a third party's detector.
In practice
Who uses a Stable Diffusion watermark remover, and when
People who generated on a hosted site, got a corner logo on the free tier, and want a clean version of their own image. Designers taking a generated background into a layout where a brand mark in the corner is simply in the way. Artists using a custom checkpoint that keeps producing a phantom signature in the same place.
And a fair number of people who arrived here having read that Stable Diffusion watermarks everything, wanting to know whether their image is marked. If that is you, the second section of this page is the answer and you may not need the upload panel at all.