Can AI Remove Your Event Photo Watermark?

Why watermarks are under pressure

Photographers and organisers who sell event photos rely on the free preview being worth less than the paid file. AI image tools have made that harder in three ways.

• General AI image editors. In March 2025, TechCrunch reported that people were using Google's Gemini 2.0 Flash to remove watermarks from stock photos. The same report noted that it struggled with semi-transparent watermarks and watermarks covering large parts of an image.
• Free one-click removers. Open-source projects such as WatermarkRemover-AI pair a vision model that finds the watermark (Microsoft's Florence-2) with an inpainting model that fills the gap (LaMa).
• Open image-editing models. On photo-editing forums, people recommend instruction-following editors such as Qwen-Image-Edit: you upload the photo and type "remove the watermark".

All three do best on the most common watermark: one logo or line of text in a corner. That area is small, so the tool only has to invent a small patch of background.

What the research says

Two papers shape how we think about this.

• Consistency is the weak point. In "On the Effectiveness of Visible Watermarks" (CVPR 2017), Tali Dekel, Michael Rubinstein, Ce Liu and William T. Freeman of Google Research showed that when the same watermark sits in the same place across a collection of photos, it can be estimated and removed from all of them at once. Their August 2017 summary explains that small random changes to the watermark on each photo make clean removal far harder.
• Large watermarks are still hard. In "Bridging Knowledge Gap Between Image Inpainting and Large-Area Visible Watermark Removal" (AAAI 2025), Yicheng Leng, Chaowei Fang and colleagues found that removal models still struggle with large-area watermarks and depend heavily on an accurate map of where the watermark is.

Put simply: a watermark resists removal when it covers much of the photo, crosses the important parts, and is hard to separate from the picture underneath.

How 9Pic AI builds its copyright overlay

In the 9Pic AI watermark editor, Create text overlay turns a line of text, such as your studio or event name, into a watermark that covers every photo. It is generated in a landscape and a portrait version, and each choice follows the research above.

• Repeated across the whole frame. The text tiles over the entire photo instead of sitting in a corner, so there is no small area to crop or fill.
• Diagonal direction. Diagonal lines of text cross faces, bodies and bib numbers, the parts a buyer cares about.
• Alternating text style. Filled and hollow letters alternate, with light ink on a dark edge, so the text stays visible on bright sky and dark clothing alike.
• Density control. Denser repeats leave less of the photo free of text.
• A faint diamond lattice. A light pattern behind the text adds a second structure that a removal tool has to deal with.

The dialog also adds a © notice and offers Recommended settings: © on, diagonal, medium size, alternating style and dense spacing. The step-by-step setup is in how to protect photos for sale from AI watermark removal.

Every overlay is also varied, so no two stamps on a photo look exactly alike. Why that matters is below.

Our removal test

We wanted to know how the overlay holds up against the tools people actually use, so we tried to remove it ourselves. We used three photos from our own library: a runner at the finish line, a runner receiving a medal, and a crowded wedding dance floor. Each got three watermarks:

• a single "© EVENT PHOTOS 2026" line in the bottom-right corner, the kind many studios use;
• the 9Pic overlay with the Recommended settings (© on, diagonal, medium, alternating, dense);
• the 9Pic overlay with weaker settings (horizontal, sparse, filled).

Then we tried five ways of removing them and checked every result at full size. We also measured how much of the text was left, and how closely each face still matched the original using a standard face-recognition model.

What we saw in each case:

• Claude declined 8 of the 9 requests outright, pointing to the © notice and suggesting the person buy the photo or contact the photographer. In the ninth, where we said we were the photographer, it said it had no tool to edit the image and pointed to the original files instead. It did not help remove the text in any reply, including when asked to write a removal script.
• The one-click remover re-creates the image path of WatermarkRemover-AI with its default settings. It cleaned the corner watermark off every photo almost perfectly. On the 9Pic overlay it found only a small patch, under 6% of the photo, and changed almost nothing: 96 to 100% of the text difference was still there.
• Estimating the outline from three photos is the attack Dekel et al. described: text that sits in the same place on every photo lines up when you compare several of them. Three photos from the same gallery were enough to map about three quarters of the lettering. Filling that map removed most of the text, but it also smeared faces. Face similarity to the original dropped to an average of 0.56 to 0.83 per photo, against 0.90 to 0.99 with the watermark still on. On the wedding photo one guest's face no longer matched at all.
• The exact outline is something an attacker does not normally have. Even then, 33 to 59% of the text difference remained as faint smudges, most visible on the busy wedding photo.
• Qwen-Image-Edit, an open image-editing model from Alibaba's Qwen team, left no visible text in the one run we could make. It did this by generating a new image: the output came back at a slightly different size, skin texture was smoothed, and an earring disappeared. The face still matched the original (similarity 0.93).

What we could not test. Gemini's image models (our account had no credit left), Grok and ChatGPT were not part of this round. In March 2025, TechCrunch reported that Claude 3.7 Sonnet and GPT-4o refused to remove watermarks. We had a single Qwen-Image-Edit run because the public demo's free daily limit ran out, so treat that result as one example, not a rate. We publish the method so others can repeat it.

What we changed after the test

After the first round we went looking for stronger attacks, and found one that worked on a single photo. Our overlay's letters were near-white with a dark edge, always in the same exact colours. An attack that finds those colours can map the letters in one photo and paint them out. On the plain overlay it left only 22% of the text difference, about as well as an attacker who knows every letter pixel.

We then tested 34 watermark designs against seven removal methods on photos from our own library. Bigger text, denser text and text in two directions did not help. What did help was taking away anything an attack could lock onto:

• Varied stamps. No two repeats of the text sit exactly alike, so there is no repeating pattern to find.
• Varied ink. The letters change colour across the photo, so there is no single ink colour to key out.
• A stronger background pattern. A tool that paints out the whole watermark has to repaint much more of the photo.
• Every overlay is unique. Nobody can reproduce yours exactly from the same text.

Every copyright overlay 9Pic generates works this way. We measured the shipped version on five sports and conference photos, three variations each. Scores are the share of the text left after the attack:

The free preview stays about as sharp as before, and faces about as recognisable. The last row is the limit we have not solved: every photo in an event carries the same overlay, so someone who combines about five of them can still strip most of the text. That is one more reason to keep the free view at 300px. A saved watermark never changes by itself, so your published photos stay as you approved them; the setup guide shows how to refresh it. We have not yet run Qwen-Image-Edit or other generative editors against the varied overlay.

What this means if you sell photos

The overlay does its job against the tools most people reach for. Assistants that respect a © notice decline, and one-click removers built for corner logos do not touch it. A corner watermark gave no protection against the same remover.

The overlay does not stop a determined person with the right tools. They can remove most of it, though in our test that often meant a smeared or redrawn photo. So treat the overlay as one layer and add a second:

• Keep the free view small. With a 300px preview, even a perfect removal leaves a 300px image, which is no use as a print or profile photo.
• Deliver full size only after payment. On events that store original photos, buyers get the clean original once they pay. The watermarked file is never the product.
• Use the © notice. It makes ownership clear to people and to AI assistants.

FAQ

Does any watermark stop every AI tool?

No. Our test shows a strong overlay stops some tools and slows or spoils others, but a capable image editor can still paint text out. That is why we pair the overlay with a small preview.

Is removing a watermark legal?

TechCrunch's report quotes law firms saying that removing a watermark without the owner's consent is illegal under US copyright law, outside rare exceptions. The rules differ by country. A © notice makes the ownership clear.

Do buyers get a watermarked photo?

No. On events that store original photos, buyers receive the original at full size, without the watermark, after they pay.

Which tools did you test?

Claude Opus 5.5, a re-creation of the open-source WatermarkRemover-AI pipeline (Florence-2 and LaMa), LaMa inpainting with outlines estimated by colour, from a gallery of photos or known exactly, and Qwen-Image-Edit on the plain overlay. Gemini, Grok and ChatGPT were not part of these rounds.

Does my saved watermark change when 9Pic improves the overlay?

No. A saved watermark stays exactly as it is, so your photos never change without you knowing. To get the latest protection, open Change text and style, select Generate, save, then run Re-Process Watermark.

What to do next

1. For events that sell photos, use the copyright overlay instead of a corner logo.
2. Keep the free large view at 300px, so the preview is never the product.
3. Follow the setup guide for your next event, or read more about 9Pic Stamp watermarking.