Most modern watermark removers (ours included) ship with two tools: auto-detect, which finds the watermark for you, and a brush, which lets you paint the mask yourself. People tend to overuse one and underuse the other. Here's the field guide.
Auto-detect: when to use it
Auto-detect is right for the common case — a clear, well-formed watermark on standard content. It's built on a model trained to recognize watermark patterns: corner logos, diagonal text overlays, repeated stamps, AI-platform marks like Sora's. When it works, it's seconds of work for a result you couldn't paint by hand any cleaner.
Use auto-detect when:
- The watermark is one of the standard kinds (corner logo, diagonal Shutterstock-style, AI generator stamp).
- It's clearly visible and not heavily overlapping the subject.
- You're working on a batch — auto-detect is the only way to scale.
- You want a quick result and the image isn't high-stakes (social post, internal review).
Brush: when to use it
The brush is the right tool when auto-detect can't see what you're trying to remove, or when you need control over exactly what gets reconstructed. It's more honest because you're telling the model exactly what you want — no guessing.
Use the brush when:
- The mark isn't a standard watermark (handwritten signature, custom logo, ad-hoc text).
- Auto-detect missed part of the watermark, or grabbed extra content you wanted to keep.
- You're removing something other than a watermark — a stamp, a date code, a stray object.
- The image is high-stakes (hero shot, client deliverable) and you want frame-accurate control.
When to use both
For most non-trivial work, the answer is "auto-detect, then brush." Run auto-detect to handle the bulk of the watermark in a couple of seconds, then use the brush to extend the mask over any edges the detector missed or to add coverage for secondary marks (corner labels, edge text, embedded stamps). Two passes, total time still under a minute.
The mental model
Think of auto-detect as recognition and brush as declaration. Auto-detect is the model saying "I think this is the watermark." Brush is you saying "this is what I want gone." When the model is right, recognition is faster. When the model is uncertain or wrong, declaration is more reliable.
Common mistakes
1. Trusting auto-detect on edge cases
On stylized or unusual watermarks, auto-detect can produce a mask that looks reasonable but misses an edge. The result will have a faint ghost of the original watermark that you won't notice until you zoom in. Always review at 100% zoom.
2. Brushing too tightly
New users often paint the brush exactly to the watermark's outline, with no margin. That tends to leave faint residue at the edges where alpha-blended pixels weren't fully covered. Add a small buffer (3-5 pixels) around the watermark for a cleaner result.
3. Brushing too loosely
On the other end, painting a giant rectangle over the whole watermark area means the model has to reconstruct a much larger region than necessary, with less context. Tight but not too tight — close to the watermark's outline, with a small buffer.
4. Forgetting that the brush works on detail too
The brush isn't just for missed watermark edges — it's also for any mark in the image you want gone. Date codes from a phone camera, accidental text overlays, a stray object in the background, a small flaw in a stock photo. Same tool, different use cases.
How I personally use them
My default workflow: drop the file in, run auto-detect, look at the proposed mask. If it looks right, run the inpaint. If it's missing a piece, I extend it with the brush and re-run. If the watermark is unusual or auto-detect produces something obviously wrong, I skip auto-detect and go straight to brush. About 80% of clips finish on auto-detect alone; the other 20% take a brush pass.
Try it
Both tools live in the same editor, so there's no commitment. Open the editor, drop in an image or video, and try the auto-detect first. If it doesn't catch what you want, switch to brush. The result preview is free — you only pay if you're happy with the export.