Your image came back with a watermark you never added
Hallucinated stock marks come from training data, not the platform. Prompt changes that suppress them, plus a clean removal pass when they appear.
A translucent diagonal wordmark across a lifestyle frame, or a small glyph in the bottom-right corner you did not put there, is almost never the platform stamping your file. It is the model reproducing a pattern it saw on the photographs it was trained on. Stock-agency images are watermarked so preview files cannot be used as finals. Enough of those previews landed in training data that "professional photograph" and "faint agency mark" got learned as the same look.
Versely does not add a visible watermark on any plan. If a mark is in the pixels, the model drew it. That is good news for the repair: a drawn mark is a region, and a region can be removed. It is bad news for the "I'll crop the corner" plan, because a hallucinated diagonal does not live in a corner.
Three marks, three jobs
A hallucinated visible mark. A logo, wordmark, URL, or corner glyph the model invented because that texture co-occurs with "photograph" in training. This is the subject of this post. It is in the pixels. It was not requested. It is not a provenance signal.
A platform or plan stamp. A badge composited by a tool or a social app at export or download. The three policies filed under "watermark" cover that case. Versely does not apply one. A social-app in-app save often does.
An invisible provenance signal. A pattern woven into the pixels or the file at generation time, readable by a detector, not by an eye. SynthID is the documented example. You will not see it, you cannot inpaint it, and it is a property of the model provider, not a toggle on your account.
If you can see the mark, it is not SynthID. If it showed up on a generation you ran here and you did not download the file through a social app first, it is not a platform stamp. Treat it as a hallucination and go to the prompt or the inpaint.
The hallucination is not folklore. In Getty Images v Stability AI, the UK High Court (judgment 4 November 2025) found historic, limited trade-mark infringement for some early Stable Diffusion outputs that reproduced Getty watermarks, and rejected the claim that the model weights themselves were an infringing copy. Getty dropped its training and output copyright claims mid-trial; those were not decided. Pending cases elsewhere are not that holding, and this post is not a rights opinion. The production fact is narrower: reproducing a stock-agency mark is a documented model behaviour, not a glitch on one endpoint.
Stop inviting the mark in the prompt
The model is not trying to brand your file. It is completing a "stock photograph" pattern. Anything in the prompt that says "stock" pulls that pattern toward the surface.
Do not write the agency look. "Stock photo," "Getty style," "editorial photograph," "shutterstock," "istock," "preview image," "watermarked." You do not need those words for a clean product shot. They are how you ask for the mark.
Do not write "no watermark." Negation is weakly modelled. Naming the concept activates it. The same mechanism that makes "no subtitles" burn in subtitles makes "no watermark" a good way to get a watermark. On flow-matching image models the negative prompt field may do nothing at all; what to write instead of a negative prompt is the general case. Convert the exclusion into a description of the frame you want.
Do write the clean plate. Empty corners. No overlay graphics. No logos that are not the product's. A finished photograph, not a preview.
| Weak | Strong |
|---|---|
| "stock photo of a woman holding a bottle, no watermark" | "a finished photograph of a woman holding a bottle, empty corners, no overlay graphics, no logos other than the bottle's own label" |
| "editorial lifestyle image, Getty style lighting" | "window-lit lifestyle photograph, clean frame, unbranded background" |
| "professional product shot" only | "studio product photograph on seamless paper, full-bleed, nothing in the corners" |
On models that still expose a working negative channel, put the nouns in that field, not the instruction: watermark, logo, signature, stamp, url, copyright. Not "no watermark." The words "no" and "don't" are not operators. They are more tokens in a string. Bare nouns are what a working negative channel can actually steer away from.
If a mark still appears, do not add five more exclusions and roll the whole picture again. The composition is probably fine. Remove the mark.
Remove the mark as a region, not as a crop
A hallucinated mark is an inpaint job. You are not "taking a watermark off." You are regenerating the pixels the mark is sitting on, from the surrounding photograph.
- Identify the shape. A corner glyph is a small, opaque stamp. A diagonal agency wordmark is a translucent band across the subject. A URL along the bottom is a line of type. The mask has to cover the whole of what you can see, including the faint parts. Zoom in. Semi-transparent diagonals are easy to under-mask.
- Mask the mark plus a margin of the photograph it sits on. Fifteen to thirty pixels of the real surface around it, feathered. Include any shadow or highlight the mark is faking. A tight mask leaves a ghost of the wordmark in the join.
- Prompt the replacement, not the removal. The pitfall on every inpaint is naming the thing you want gone. Describe the surface that should be there: skin, brick, seamless paper, sky, product label. Restate the light.
A removal prompt that usually lands:
Continuation of the existing photograph: clean skin and hair in this region, window light from camera left, no overlay, no text, no logo.
For a corner glyph on a studio sweep:
Seamless light-grey studio paper continuing across this corner, even studio lighting, empty, no objects, no text.
Drive that from the photo editor. Keep denoising in the middle of the range so the surrounding texture is the constraint. Too high and you get a new patch of photograph that does not match the grade. Too low and the wordmark survives as a stain.
Two or three tries. If a diagonal band still ghosts through, widen the mask; you under-painted the translucent part. If the subject under a diagonal (a face, a label) starts to drift, you over-painted. Split the job: remove the band in two or three smaller masks rather than one mask across a face.
What not to do:
- Do not crop the corner and call it done. A corner crop changes the composition, wrecks the aspect ratio the placement wants, and does nothing to a diagonal. It is also the move the watermark glossary already flags as a pitfall, because it does not touch anything embedded, and it does not touch a mark that is not in the corner.
- Do not blur or smudge the band. A soft rectangle over a diagonal is legible as a patch. On a paid placement it looks like you hid something.
- Do not run an upscaler to "clean it up." An upscaler reconstructs detail from what is in the frame. A burned-in mark is detail. You will get a sharper watermark.
- Do not re-roll the image unless the mark is everywhere. A full regeneration is how you lose the frame you liked to chase a region you can inpaint.
After the removal, look at 100 percent. Translucent diagonals fail in the mid-tones. If you still see the ghost, you are not done.
Prevent it on the next batch
Once you have seen the mark on a model, write the clean-plate language into the standing prompt for that job. A single extra clause ("empty corners, no overlay graphics") is cheaper than a removal pass on forty frames. If a particular model keeps stamping a diagonal anyway, switch model rather than stacking negatives. The prior is in the weights.
Removing a hallucinated agency mark does not discharge a synthetic-media disclosure duty, and leaving one in does not satisfy one. The mark is a texture. The duty is about how you present the content.
FAQ
Did Versely put this watermark on my image?
No. There is no visible watermark on any plan. A mark you can see in a generation that never left this pipeline was drawn by the model, from a training-data pattern, not composited by the platform. Remove it as a region, or change the prompt so the next sample does not include it.
Is a hallucinated Getty-style mark the same as SynthID?
No. SynthID is an invisible provenance signal added at generation by a provider, readable by a detector. You will not see it in the corner. A diagonal wordmark or a corner glyph is a visible hallucination. Cropping, inpainting, or re-rolling treats only the visible one. The SynthID breakdown is the other mechanism.
Will putting "no watermark" in the prompt stop it?
It often does the opposite. The tokens for the mark enter the conditioning, and negation is a weak push. Describe the clean frame instead, and if the model has a working negative field, list the nouns (watermark, logo, signature, stamp) without "no." If the field is a no-op on that model, skip it.
Is it a legal problem if the model drew a stock-agency wordmark?
It can be, which is one more reason to remove it rather than publish it. The UK High Court in Getty Images v Stability AI (4 November 2025) found historic, limited trade-mark infringement on some early-version outputs that reproduced Getty watermarks, and rejected the secondary copyright claim against the weights. That is a specific holding in a specific jurisdiction, not a global rule, and it is not advice about your campaign. Operationally: do not ship a frame that bears another company's mark, even one the model invented. Inpaint it.