How to remove text from an image

Removing text from a still is inpainting with a smaller story. You mark the letters. A model fills the hole from the surrounding pixels. Lumenfix remove text is that job and nothing else. This guide covers the mask, the two quality tiers, and the line we will not cross.

Only text you are allowed to remove

If you added the caption, the sticker, the timestamp, or a draft watermark on your own photo, this page is for you. If a photographer, an agency, or a stock site put a mark on the file, stop. Lumenfix does not ship a watermark remover and will not add one. Detection of C2PA stays read-only on the AI image detector.

The same engine powers remove person and remove object. The difference is copy and the default story, not a secret second model.

Paint a mask that matches the photo

The brush writes a PNG with the same width and height as the upload. Transparent pixels are the region to fill. That is the contract the server sends to LaMa and to the precision erase path. A mask that is a different size will fail. A mask that only covers half a letter will leave a ghost.

Use a brush a bit larger than the stroke. Text has antialiased edges; if you hug the glyphs too tightly, the model keeps a faint halo. If you paint a third of the frame, you are asking the model to invent a wall. That is when identity and architecture start to drift.

Undo and clear exist because everyone over-paints. There is no automatic “find all text” button on purpose. OCR-plus-erase is a different product and a good way to delete a shop sign you meant to keep.

Free LaMa versus Pro precision

The free path is LaMa on Replicate. It is cheap, deterministic enough for signs and captions, and it does not try to be an art director. Three free jobs per UTC day on this tool. No watermark on remove — the hole fill is the deliverable.

Pro can switch to the AI precision erase, which is gpt-image-2.5 via apimart with the same mask. Use it when LaMa leaves a smear on a repeating texture: tiles, brick, a knitted sweater. Do not use it because you want the model to write a different word. If the letters change, that is a failure. Delete the output.

When the site-wide free budget is spent, anonymous jobs return busy. Pro is unaffected. If apimart fails, the error is provider_error. A free quota tick is already spent; a credit debit would roll back. Remove does not spend credits today.

A workflow that does not make things worse

1. Work on a copy of the original, not the only scan of a print.
2. If the file is tiny, upscale after the erase, not before, unless the letters are unreadable at the current size. Upscaling junk makes more junk.
3. If the photo is soft, enhance after the hole is gone. GFPGAN on leftover letter fragments looks like a rash.
4. If you also need more canvas, extend after the text is gone so the model does not clone the caption into the new border.
5. Check the result at 100%. Look at the area around the old letters. That is where LaMa usually fails, not in the exact center of the stroke.

What will look fake

Large text over a face. The model has to invent skin. That is how you get a third eyebrow. Large text over a grid of windows. Perspective breaks. Stylized posters where the type is the design — you cannot remove the word without removing the poster.

Blurred-but-present letters belong on unblur text, not here. Unblur is a deblur pass that is not allowed to guess a new sentence. Remove is the opposite: the sentence is supposed to disappear.

Browser versus server

Unlike the detector and PNG to SVG, remove uploads the still and the mask. The pixels have to reach LaMa. Object storage keeps them under tools/ and expires them. Do not upload a file you cannot store for a day.

There is no prompt box on the free path. The precision path has a locked “fill the background naturally” instruction. We will not add a box that says “put a different slogan here.” That would be generation.

After the text is gone

Download immediately. If you need a print, upscale. If you need a different aspect ratio, extend. If you are unsure whether the rest of the photo was generated, run the detector on a copy that still has the text — inpainting can itself look generated in the hole.

If you work in a newsroom or an archive, write down what you removed. A later reader should know the caption was yours and not part of the original glass plate. The downloaded file will not carry that story. A sidecar text file is enough.

Batch work is a Pro conversation about rate limits, not a zip uploader. The Worker still wants one still and one mask. If you have forty slides with the same lower-third, you will paint forty masks. That is slower than a dedicated broadcast tool and cheaper than commissioning a desk to rebuild each frame.

Phones introduce a second failure mode: the text is not pixels in the photo but pixels in a screenshot of a chat. The background is a UI gradient that LaMa will happily extend into a mushy rectangle. Crop to the photo first, or you are inpainting a Messages bubble.

When you are done, glance at the detector only if the rest of the still is in dispute. The filled hole may score as generated even when the rest of the frame is a camera original. That is expected. Do not use the hole as proof the whole picture is fake.

Color mismatch after a fill is common on sunset skies and mixed fluorescent kitchens. If the patch is a slightly different white, run a second, smaller brush on just the seam instead of repainting the whole word. A second cheap LaMa pass on a thin crack is better than one huge precision erase that rewrites a brick wall.

If you are deciding between the three remove URLs, pick the one that matches the searcher you are helping. The engine is the same. The FAQ examples are not. A “remove person” visitor should not land on a page that only talks about captions.

Removing text is a small, old computer-vision job. Treat it that way. Paint the letters you added, fill the hole, and stop before the model starts writing a new picture.

Keep a note of the mask you used if the file will be published. A later editor should be able to see that the caption was yours and that the fill is inpainting, not a second camera exposure. That sentence is the whole ethics of this tool.

Free jobs share a daily counter across remove-text, remove-person, and remove-object because they are one tool id. Three holes a day is the trial. If you need a fourth, wait for the next UTC day or use Pro. Credits are not spent on LaMa. Precision erase is the paid model, and a provider failure rolls those credits back.