How to Remove Text from an Image
Updated Aug 28, 2026 · 6 min read
What text removal can and cannot do
Text on an image is just pixels covering other pixels. Removing it means detecting the letters and then reconstructing whatever was underneath. The Text Remover does both: an OCR-style detector finds text regions, and an inpainting model fills them from the surrounding context.
It works best on:
- Captions and meme text over photos, where the background is a natural scene.
- Date stamps from old digital cameras, usually orange digits in a corner.
- Burned-in subtitles on video frames and screenshots.
- Your own text overlays on marketing images you want to reuse with new copy.
- Placeholder text on templates and mockups.
It struggles when the text covers something detailed and unique, such as letters across a face or a product label, because there is nothing for the model to infer the hidden detail from.
A note on what you should not use it for: removing another creator's watermark, credit line or copyright notice from an image you do not own. Beyond being unfair, stripping copyright management information can carry legal liability in many jurisdictions. Use this tool on images you created, own or are licensed to edit.
Step 1: Try Auto detect first
Upload your image to the Text Remover and leave the mode on Auto detect. The detector highlights every text region it finds with a red box; you can untick any box you want to keep, such as a brand name on a product that is supposed to stay.
Auto detect handles Latin, Chinese, Japanese, Korean, Arabic and Cyrillic scripts, printed or bold display fonts, and text at moderate angles. It occasionally misses handwriting, very thin or very low-contrast text, and stylised logotype lettering.
Click Remove. Processing takes about three to five seconds and costs 1 credit for the HD download; the preview is free. Inspect at 100 %: the filled areas should be indistinguishable from their surroundings, with no faint letter shapes and no smudged rectangle.
Step 2: Use the manual brush for precision
Switch to Manual brush when auto detect misses something, removes too much, or when the text sits on a complex background where you want to control exactly which pixels are regenerated.
How to paint a good mask:
- Set the brush size a little larger than the stroke width of the letters. For a 24 px font, a 30 to 36 px brush is right.
- Paint over the letters with a small margin. You do not need to trace each glyph; a single sweep across the line of text is fine.
- Include the drop shadow or outline if the text has one. A removed word that leaves its shadow behind is the classic giveaway.
- Use the eraser to pull the mask back from anything you want preserved, for example a face right next to a caption.
- Run, inspect, and if one word remains, paint just that word and run again. Each run is 1 credit.
For text over a repeating texture like fabric, brick or grass, keep your strokes tight. The more of the pattern that survives around the mask, the better the model can continue it.
Step 3: Fix difficult backgrounds
Text across a face or a product. After removal, the model will have guessed at the covered part. If the guess is poor, switch to the Object Remover and paint a slightly larger region that includes natural boundaries (the whole cheek, the whole label), giving the model a cleaner area to reconstruct.
Text on a gradient sky. Usually perfect, but banding can appear if the source is a heavily compressed JPEG. Run the result through the Photo Enhancer to smooth it.
Text along an edge, such as a caption in the bottom bar. If the bar itself is a solid colour, the fill will be flawless. If the bar was added by an app and you would rather remove the whole bar, crop it out instead; cropping is free and always cleaner than inpainting.
Semi-transparent text. The detector handles it, but paint the mask a little wider than you think, because the faint edges of transparent letters extend further than they appear.
Very large text covering more than about a third of the image. Inpainting quality drops with area. Consider whether the image is still usable, or whether a version without the text exists.
Batch and video frames
If you have a folder of photos with the same date stamp in the same corner, the manual mask can be saved and reused. Paint it once, save it as a preset, then apply it to every image in the batch. This is faster and more consistent than auto detect for stamps, because the position never changes.
For subtitles on video, export the frames you need as PNG, process them with the same saved mask, and reassemble. Full video text removal is not yet supported; the Video Background Remover isolates a subject rather than erasing overlays.
Pricing recap: 1 credit per image in either mode. On Starter ($9.99 for 250 credits) that is about $0.04 per image; on Pro it is about $0.03.
Frequently asked questions
Can it remove text from a screenshot of a chat?+
Yes, and flat UI backgrounds fill perfectly. Untick any message bubbles you want to keep before running.
Will it remove a watermark from a stock photo?+
Please do not use it for that. The tool is intended for images you own or are licensed to edit. Removing a copyright notice from someone else's work can be unlawful.
Does it support Chinese and Japanese text?+
Yes. Auto detect recognises CJK scripts as well as Latin, Arabic and Cyrillic.
What if a faint outline of the text remains?+
The mask was too tight. Switch to manual brush, paint a few pixels wider than the letters including any shadow, and run again.