Selection first, then reconstruction
Clean a flattened caption, an expired promo line, or a date that no longer belongs. You mark the text, the model rebuilds the surface behind it, and the download keeps the original pixels everywhere you did not paint.
Upload the photo with baked-in text
JPG, PNG or WebP under 20MB. Each side must be between 256 and 4096px.
This run costs
20 credits
Your balance
--
A percent-off tag or a currency line is baked into the packshot. Brush the letters and the surface under them only, so the label, the logo, and the product itself stay exactly where the photographer placed them.
A poster or flyer carries a date that has passed. Mark that single line, leave the venue name and the artwork untouched, and reuse the layout for the next date instead of rebuilding the design.
Handwriting on fabric, wood grain, or concrete needs the reconstruction to continue the texture, not smear a flat patch. A tight mask around the ink gives the model the neighbouring pattern to follow.
Choose a JPG, PNG, or WebP file up to 20 MB, each side between 256-4096 px per side, and at most 8 megapixels. The browser prepares a working copy at the same frame so the mask lines up with the pixels you see.
Paint over every letter that should go, with a small margin past the ink for tight serifs. Erase, undo, redo, and resize the brush freely. An empty selection cannot start a paid run.
Confirm you have the right to edit the image, then start one 20 credits run. The model reconstructs the marked background only, the workbench composites it back onto your original, and you compare before you download.
Each pair keeps the same scene, camera, and crop. Only the marked text and the surface behind it differ between the two frames, because that is the only region the run is allowed to touch.
Concept example. Each pair illustrates the mask-and-reconstruct job on a photo you are allowed to edit. These are not measured customer results.
Concept example
Concept example
Concept example
Reconstruction is an informed guess about hidden pixels, not a recovered layer. Tight kerning, dense serifs, and busy patterns need a slightly larger mask than the ink itself, so a faint stroke can remain.
A remove text from image task looks like an eraser job, but it is really a reconstruction job. Once a caption is flattened onto a photo, the pixels under the letters are gone. The only honest goal is to rebuild a plausible background from the colour, light, and texture that surround the marked area, while keeping everything else exactly as it was.
IMG AI makes the selection the user’s decision instead of a detection guess. You paint the letters that should go, the model reconstructs that region, and the workbench composites the reconstruction back onto your original so untouched captions, brand marks, and grain survive even when the model drifts.
The whole edit is bounded before it runs: one JPG, PNG, or WebP upload, one marked region, 20 credits per run, and a PNG composite at your working frame. There is no layer stack to learn and no desktop install; the mask editor runs in the browser and the result lands in My Creations.
Pick the workflow that matches the file and the risk. Targeted reconstruction, a desktop healing brush, and a detect-everything wipe are different jobs with different trade-offs. Competitor cells state only what is publicly documented and stable; unverified facts say so instead of guessing.
| Axis | IMG AI Remove Text | Desktop clone stamp or heal | Automatic detect-and-wipe text |
|---|---|---|---|
| Best for | A flattened JPG, PNG, or WebP photo where you know which letters should go and which marks must stay. | Retouchers who want pixel-level sampling and layered history on a workstation. | A throwaway frame where every glyph should vanish and a leftover logo is acceptable collateral. |
| Control over the region | Brush, erase, undo, redo, and zoom until the mask is deliberate. A required mask gates the paid run. | Manual sampling, healing, and a full history stack on a local file. | One-click detection with little regional choice. |
| What happens to unmarked pixels | Kept from your original. The workbench composites the reconstruction inside the mask only. | Untouched unless you paint them; the operator decides every change. | Not verified here |
| Cost of a single edit | 20 credits, shown on the button before you run it. | Software licence plus the operator’s time; the per-edit price is not verified here. | Not verified here |
| Where it runs | In the browser at imgai.pro, no install. | In a desktop photo editor on a workstation. | In a browser or app that ships its own detector. |
| What it refuses to do | Read the type, replace it, or run with an empty mask. Identity documents and copyright watermarks are out of scope. | Nothing by itself; it is a manual tool whose quality depends on the operator. | Distinguish the one caption you want gone from the brand mark beside it. |
Every run reserves 20 credits on your IMG AI account, and the button shows that number before you press it. The quote is fixed per run; it does not change with image size or how many lines you mark in one mask.
The upload can be JPG, PNG, or WebP up to 20 MB, with each side between 256-4096 px per side and at most 8 megapixels. The browser prepares a working copy at that frame, and the download is a PNG composite at the same working size.
Finished runs are saved to My Creations, so a result can be opened and downloaded again later without starting another paid task.
It means marking flattened letters on a JPG, PNG, or WebP photo you may edit, reconstructing the surface behind that mark, and downloading a composite that keeps every unselected pixel from the upload.
No. You box or brush the letters. An empty mask cannot start a paid run, and that is how nearby captions and logos stay put.
For accounts with an active paid entitlement, the workbench composites the reconstruction only inside the mask and keeps the original pixels everywhere else, including the original width and height. Accounts without an active plan receive the watermarked provider preview instead. Compare the split view before you download.
Not on this page. This tool produces a clean background where the text was. Add replacement type in a design tool, or use Generative Fill when the job is a described change.
No. Copyright watermark removal and identity documents are out of scope. Never upload passports, licences, or other identity documents.
One signed-in run costs 20 credits. Recovering the same request id after a dropped connection does not charge again. A new mask on a new request id is a new charge.
