Remove text from images you own or are allowed to edit. Erase captions, promo copy, dates, and labels by boxing or brushing the words, keep the rest of the frame untouched, and let the tool rebuild only the selected background.
Source image / 01
Upload the picture that contains the text you want gone.
JPG, PNG, or WebP · 10 MB max · 128–4,096 px per side
Canvas / mask preview
100%Your mask canvas appears here
Upload an image to box or brush over the text, then generate a clean rebuild.
Text removal studies
Each study shows the same decision: the words on the image are the target, and everything around them—product, face, texture, and composition—is a preservation goal. Results are reconstructions, not copies of hidden pixels.
Case 01 · Poster / caption
A short punch line sits over a gradient corner. The box selection covers only those words so the layout, headline, and imagery keep their original pixels.
ReviewCheck the repaired corner for banding, grain, and a continuous gradient before republishing.
Case 02 · Product / promo
Discount copy is baked over a clean studio surface. Masking just the letters lets the tool continue the smooth backdrop and shadow instead of regenerating the product.
ReviewCompare the label, reflections, and shadow direction against the source at 100%.
Case 03 · Photo / timestamp
A bright date stamp overlaps foliage and brick. A tighter brush selection follows the numerals so nearby leaves and mortar lines are not asked to change.
ReviewZoom into the texture where the digits were; repeating patterns are the hardest area to rebuild.
Definition
Removing text from an image means painting a mask over the words you want gone, then rebuilding the pixels underneath so the background looks continuous. It is an inpainting job: the tool only edits the masked region and leaves the rest of the picture alone.
There are two different jobs people describe with the same phrase. The first is editing text that is still live—retouching a design file, swapping a headline, or fixing a typo. The second is cleaning text that is already flattened into pixels, such as a caption exported into a JPEG, a promotion baked over a packshot, or a timestamp burned in by a camera. This page handles the second job and links to editing tools for the first.
Because a flattened image no longer stores the pixels hidden behind the words, no tool can recover the exact original background. The realistic goal is a plausible reconstruction: match the surrounding colour, texture, lighting, and perspective so the repaired area stops drawing attention. Mask-guided inpainting makes that goal more controllable than a prompt-only edit, because you decide precisely which pixels are allowed to change.
The workbench therefore centers on selection rather than on typing. A rectangle is fast for a headline block or a caption strip, while a soft brush follows curved letters, small digits, or text that crosses several textures. Multi-region selection and undo mean one task can clean a date stamp, a corner caption, and a promo line without starting over.
Practical workflow
The sequence is short because the important decision—what may change—is made visually in the workbench before anything is submitted.
Choose a JPG, PNG, or WebP file up to 10 MB. Each side must be between 128 and 4,096 pixels, JPEG is capped at 1 megapixel, and PNG or WebP at 4 megapixels. The page checks the decoded image, not just the extension.
Work from the highest-quality copy you have; heavy JPEG artifacts make the rebuild harder.
Drag a rectangle around straight text or switch to the brush for curved and angled words. Add every region that should be cleaned, zoom in for small print, and undo any selection that grabs more than the letters.
Confirm the mask covers the full glyphs and nothing that must be preserved, such as a logo or product label.
Press Generate to rebuild the masked regions. Each run costs 30 site credits and produces one image. Compare the result with the original using the slider, then download the finished PNG or retry with a tighter mask.
Inspect the repaired area at 100% and confirm no text stroke survives and no unselected content changed.
Selection toolkit
Selection is the real skill on this page. The tighter the mask, the less the model has to guess, and the more of your original image survives untouched.
Drag a box around a block of text such as a caption, a headline, or a promo banner. Fast, predictable, and easy to enlarge slightly when letters have a soft edge or shadow.
Best forCaptions, price tags, banner copy, and straight text lines.
Paint over irregular or curved words, digits, and watermarks with an adjustable round brush. Brush strokes are stored as smooth paths so they scale back to the full-resolution mask.
Best forDate stamps, cursive script, angled labels, and text over texture.
Add as many boxes or strokes as the image needs, then remove the last action with Undo or clear everything and start again. Each region is part of one mask and one task.
Best forImages with several separate text areas, like a poster with a caption and a code.
Zoom the canvas to place the mask precisely on small letters. Selection coordinates are stored in image space, so the exported mask always lines up with the uploaded pixels.
Best forSmall print, thin fonts, and high-resolution photographs.
Cover the full glyph, including anti-aliased edges and any drop shadow. A mask that stops halfway through a letter often leaves a visible stroke behind.
Method selection
Different cleanup methods change different amounts of the picture. Choosing the right one is mostly about how much of the original image you need to keep.
| Method | What changes | Control you keep | Best for | Main risk |
|---|---|---|---|---|
| Mask-guided inpainting | Only the masked region is rebuilt | You draw the exact region with boxes and brush strokes | Captions, promo text, and date stamps on images you must preserve | A poor mask leaves a halo or misses part of a glyph |
| Prompt-only regeneration | The whole frame may be reinterpreted | You describe the change in words and hope the model localizes it | Creative restyling where exact preservation is not required | Unrelated faces, labels, or textures can drift |
| Clone or heal by hand | A manually sampled patch is copied over the text | You place every sample and blend by hand | Tiny marks where a human editor wants frame-by-frame control | Repeating patterns and gradients are hard to match by eye |
No method recreates the true hidden pixels of a flattened image. Mask-guided inpainting reduces unintended changes; it does not guarantee a perfect reconstruction.
Use cases
Most legitimate jobs share one condition: you own the image, license it for this purpose, or have the rights holder’s permission to modify it.
Is an old event date, price, or discount still baked into a reusable visual?
Try this: Box the old copy tightly, rebuild the background, then add the current wording in a live editor so future updates stay editable.
ReviewCheck that the new background matches the original before you place fresh text on top.
Does a caption or subtitle sit on a frame you want to reuse without the words?
Try this: Use a slightly padded rectangle when the caption has a shadow, and keep the mask away from the subject’s face and edges.
ReviewLook for a faint outline where the caption used to be, especially over gradients and skies.
Do you need a label-free or copy-free version of a packshot, mockup, or poster for reuse?
Try this: Brush tightly around small print, and separate each text block into its own region so one bad area is easy to undo.
ReviewVerify packaging text, logos, and straight edges that must stay identical to the source.
Commercial comparison
These services overlap, but they make different trade-offs between automation and control. The comparison focuses on the decision that matters here: how precisely you can choose the text region and how much of the original image survives. Competitor names appear for clarity; this page does not send you to competitor sites.
Workflow comparison reviewed
| Decision point | Nano Banana Pro | Cleanup.pictures | Pixelcut |
|---|---|---|---|
| Selection model | Rectangle and brush selection with multiple regions, undo, clear, and canvas zoom before you submit. | A brush-first erase workflow aimed at quick object and blemish cleanup. | A general AI photo editor with cleanup tools grouped among many other features. |
| What the model is allowed to change | Only the masked region is rebuilt; the source pixels outside the mask are the preservation target. | The retouch is centered on the painted area, with behavior that depends on the current tool. | Because it is a broad editor, the final result depends on which feature and settings you choose. |
| Text-specific review loop | Zoom, undo the last region, and compare original against result with a slider before downloading. | Review happens inside the erase canvas for the painted area. | Review happens in the broader editing project alongside other adjustments. |
| Entering the workflow | The first screen is the tool itself: upload, select, and generate without a setup wizard. | Upload-and-erase flow focused on a single cleanup task. | Editor-first flow where you then choose an AI cleanup feature. |
| Cost behavior on this page | 30 site credits per generated image; selection, zoom, undo, and the comparison slider are free. | Plan, credit, and export rules depend on the current product and account. | Plan, credit, and export rules depend on the current editor and account. |
Choose Nano Banana Pro when you want to decide the exact text region, keep everything else as-is, and review a mask-guided rebuild before downloading. Choose a broader editor when the text cleanup is one small step inside a larger design job.
Operating facts
The workbench validates these boundaries before a paid task starts. They describe the implemented page, not a promise that every reconstruction will be invisible.
One generated image costs 30 site credits. Selecting, zooming, and comparing do not repeat the task.
Selection and mask drawing happen in your browser. On upload, the server checks the actual image container, size, dimensions, and content moderation result, then stores both the source and the mask under your signed-in account. Before generation, the server verifies that both stored references belong to you and that the mask matches the source dimensions.
The generation task is created through the existing workbench provider path, which reserves credits, tracks the task, handles provider failures, and refunds failed runs. Completed images are fetched through the authenticated download route, so a result is only served to the account that requested it.
Credits are checked and consumed by the server when the task starts. Changing the selection, zoom, or comparison slider after a result does not create a new charge, and a retry is a new task.
Common questions
Short answers about input requirements, what can change, how the workflow behaves, and what the result can realistically deliver.
Upload a JPG, PNG, or WebP image, then drag a rectangle around straight text or use the brush for curved and angled words. Add every region that should be cleaned, press Generate, review the result with the comparison slider, and download it. The visible steps are upload, select, generate, and download.
The tool rebuilds only the region covered by your mask, so captions, promo copy, dates, labels, and watermarks are the target. Everything outside the mask—faces, products, logos, and backgrounds—is meant to stay as-is. If a selection accidentally covers a logo, undo it before generating.
This page does not claim automatic OCR text detection. You mark the words yourself with the box or brush, which keeps you in control of exactly what changes. That manual selection is the feature, not a missing automation.
The page accepts JPG, PNG, and WebP files up to 10 MB. Each side must be between 128 and 4,096 pixels. JPEG is limited to 1 megapixel, and PNG or WebP to 4 megapixels. GIF, PDF, and video files are not accepted.
Each generated image costs 30 site credits and returns one result. Drawing, zooming, and undoing the selection are free. A retry is a separate task and is charged again, so tighten the mask first. If the provider fails, the task follows the existing refund path for the account.
The goal is a seamless rebuild, but the result is a generative reconstruction, not the original hidden pixels. Thin fonts, gradients, repeating textures, and drop shadows are the hardest cases and may leave a faint trace. Compare the whole image at 100%, and refine the mask if the rest of the photo shifted.
No. Only edit images you own, license for this purpose, or have explicit permission to modify. Removing visible text does not grant rights to the underlying work, and some text—such as copyright notices or attribution—should not be removed. This page offers workflow guidance, not legal advice.
Uploads and results are tied to your account and served through authenticated routes, so another user cannot download your output. Task state, credits, and refunds run through the existing workbench workflow. Avoid uploading sensitive material you would not want processed by an AI image service.
Return to the workbench
Start with an image you are allowed to edit, mask only the text, and rebuild the background without touching the rest of the frame. Related tools are here when you need a full editor, object removal, or a bigger canvas.