The job
What this Stable Diffusion watermark remover handles
The upload panel sends your image into our media workflow, where detection and the edit run together and a finished file comes back. Before you spend an upload on it, this page walks a triage step, because Stable Diffusion watermark covers three different situations and only one of them is an image edit.
That triage is the part competing tools skip. A commercial page we read for this term handles visible overlaid marks and does not mention invisible watermarking anywhere on it, which leaves a reader who has the other problem no way to find out.
Do this first
Triage: which Stable Diffusion watermark do you have?
Three possibilities, and you can tell them apart by looking at the file and remembering where it came from.
- You can see a logo
- A brand mark in a corner. It was added by the hosted platform you generated on, not by the model. This is the case our workflow edits.
- You can see nothing
- The reference Stable Diffusion code writes an invisible payload using the invisible-watermark library, applied by a put_watermark call in txt2image.py. Nothing you do to the visible image addresses it.
- You see a scribble that looks like a signature
- Custom checkpoints trained on signed artwork reproduce signature-shaped marks. That is a style artefact and an ordinary retouching job.
- You generated locally
- No hosted platform added anything, and the watermark step in the reference script can be switched off at source. There may be nothing to remove at all.
The run
How to remove a Stable Diffusion watermark in three steps
Once triage says you have a visible mark, the run itself is short.
- Step 1
- Keep the original file. There is no undo inside the workflow and the result comes back as a separate rendering, so you want the source to compare against.
- Step 2
- Upload one image. Detection and the edit run together, and the output is a re-rendered file rather than the original with a patch pasted over the corner.
- Step 3
- Open both at full size and inspect the region the mark covered. Flat backgrounds almost always survive cleanly. Text, hands and faces are where a regeneration shows.
The one you cannot see
The invisible Stable Diffusion watermark, and what edits it survives
The invisible-watermark library that Stable Diffusion's reference code calls publishes its own robustness results, and they are worth having in a workflow because they contradict the usual assumption. The library states its methods are not robust to resize or to aspect-ratio-changed crop, but are robust to noise, colour filter, brightness and JPEG compression.
So the routine actions people take to sanitise a file mostly do not touch it. Its published test table shows the frequency methods failing at a 50% resize and failing under rotation, and a 7 by 5 crop defeating the frequency methods while the deep learning method survives. The library also records a known defect: none of the algorithms performs well on web page screenshots or on posters with a homogeneous background colour.
Our method, stated
Regeneration, not a patch over the corner
Our workflow regenerates the image. It re-renders the file rather than peeling a marker off the original or covering it with reconstructed pixels, and the thing you download is a new rendering of your picture. Knowing that changes how you review the result, which is why the page says it rather than describing the output as the original minus a logo.
It also sets the boundary on what we will promise. No guaranteed clean result on every file, and no statement about what any decoder will report about the output afterwards, ours or a third party's. A page that promises you a specific verdict from someone else's detector is promising something it cannot know.
The check nobody else runs
Whose image is it, and what does the licence say?
Add this to the triage, because none of the competing pages does. Removing a watermark grants no right to use an image, and modification is forbidden outright by a great many licences. Where a mark is present to enforce a licence, taking it off changes the file and not the terms.
The workflow this page supports is your own output: images you generated from your own prompt, on a platform whose export mark you are entitled to remove. Stripping a third party's branding from their work is a different activity and it is not the one being described here.
Asked directly
Does every Stable Diffusion image come back clean?
No, and any workflow built on the assumption that it will needs a review step. Regeneration reconstructs what was under the mark, and reconstruction over flat colour is invisible while reconstruction over detail is a guess that sometimes shows.
Build the comparison into the process rather than hoping. Two windows, both files at full size, attention on the covered region. It takes seconds and it is the difference between shipping a clean image and shipping a smeared one.
Where it fits
Who runs images through a Stable Diffusion watermark remover
Designers pulling generated backgrounds into layouts, where a corner logo simply cannot stay. Teams that generated on a hosted free tier and need their own images without the platform's branding. Artists on custom checkpoints dealing with a recurring phantom signature.
And a large group who never upload anything. They arrived because they read that Stable Diffusion watermarks every image and they want to know whether that is true of theirs. For them the triage section is the whole answer.