Text behind image
Put a word behind the person in a photo, so they stand in front of it. The subject is cut out here, inside this tab, which is why the picture never has to leave your machine and why the file comes back at the size you dropped in.
Drop a photo in, drag the words behind the subject, save the picture.
This one is hard work for a phone. The cutout is a model running inside your browser, and a phone tab gets a small slice of memory to run it in. Expect it to be slow, and on the larger model to fail outright. For a picture that matters, open this page in Chrome on a desktop or a laptop.
Look
Layers
Text
Colour
Place
Drag the words on the picture. Arrow keys nudge one pixel, with Shift ten.
Cutout
Free · no sign-up · nothing uploaded. The cutout model runs in this tab, so the photo stays on your device. The weights are fetched once from a CDN and cached, and after that the tool works offline.
Removed from this browser.
Kept in this browser only. Create an account to keep them on every device.
How to put text behind an image
Drop a photo in
Drag one onto the panel, click to choose, or paste a screenshot. The model finds the subject and cuts it out here, in your tab.
Drag the words where you want them
A text layer is already there, sized to the picture. Drag it, nudge it with the arrow keys, and it slides behind the subject as it goes.
Save it at full resolution
Pick PNG, JPG or WebP and press Save. You get the picture back at the size you dropped in, with no watermark.
Using the text behind image editor
The picture is on the left and every setting is on the right, with nothing behind a tab. That is the one decision the whole panel rests on. The effect is three layers deep, and the reason it usually looks wrong is a legibility problem rather than a cutout problem, so the controls that fix it have to be findable rather than hidden two menus down.
What you can hand it
Drag a file onto the panel, click the frame to choose one, or press CtrlV to paste a screenshot. That last one is the shortest route on a phone, because there is no file to find first.
| Accepts | Notes |
|---|---|
| PNG, JPG, WebP, AVIF, GIF, BMP | Anything your browser can decode as an image. |
| Any file size | No megabyte cap, because there is no upload to pay for. The nearest rival stops at 10 MB. |
| Up to 40 megapixels | Past that it is refused rather than left to run the tab out of memory. |
| A font file | Drop a .ttf, .otf, .woff or .woff2 on the panel and it joins the font list. |
| One picture at a time | There is no batch here. See the limits below. |
The picture, and moving the words on it
A text layer is already in the panel before you drop anything, sized to the picture as it lands. So the effect is on screen with nothing to press, which is the point: this is an effect people want to see before they decide to build one.
| To do this | Do this |
|---|---|
| Move a layer | Drag it on the picture. A blue line appears when it reaches the exact middle, and it snaps there. |
| Move it precisely | Click the picture, then the arrow keys. One pixel a press, ten with Shift. |
| Resize it | Drag any of the four corner handles. Or use Size in the inspector. |
| Rotate it | Drag the round handle above the box. Hold Shift to step in fifteens. |
| Pick a different layer | Click its words on the picture, or its row under Layers. |
| Check the cutout | Hold Hold to see the photo under the panel. It shows the original for as long as you hold it. |
Look: the six presets
Each chip writes a complete set of appearance values onto the selected layer: the face, the weight, the tracking, the colour, the outline, the shadow and the blend mode. None of them touches your words, your position or your rotation, so pressing through all six to see which one suits the photo costs nothing.
| Preset | What it is for |
|---|---|
| Poster | Heavy, tight and white. The default look of this effect, and the one that reads at thumbnail size. |
| Outline | Almost no fill and a white stroke. For a busy photo where a solid word would fight the background. |
| Glow | White with a soft white halo. The fix for a light word over a light sky. |
| Wide | Letterspaced, with a small dark shadow. Reads as a caption rather than a headline. |
| Ghost | Half opacity on the screen blend, so the word takes the photo’s own colour. |
| Stamp | Monospaced with a dark outline. For a date, a place or a serial number. |
Text: the words and how they are set
Everything in this group applies to the one layer selected. The box at the top takes line breaks, so a two-line headline is two lines in the box.
| Control | What it does |
|---|---|
| Font | Nine families. The first three are already on this page, so choosing one costs no download at all. The rest are stacks of faces your device is likely to have. |
| Use a font file from this device | Reads a font off your disk and adds it to the list. Nothing is uploaded. It lasts as long as the tab. |
| Size | A percentage of the picture’s width, not a pixel count. So a design keeps its proportions on a photo of any size. |
| Weight | Light, Bold, Black. A face with no black cut gets a synthesised one. |
| Align | Only matters on more than one line. It aligns the lines against each other, not against the picture. |
| Spacing | Letterspacing, as a percentage of the type size. Negative tightens, which is what a heavy poster word wants. |
| Line height | The gap between lines, as a multiple of the size. 1.00 is the tight setting display type usually wants. |
Colour: the group that decides whether it is legible
This is where the effect is won or lost. A white word on a bright sky disappears, and every other free editor for this effect leaves you with nothing but the colour picker to fix it. There are four controls here and three of them exist for that one problem.
| Control | What it does | Reach for it when |
|---|---|---|
| Fill and opacity | The colour of the letters, and how solid they are. | Always. Dropping to about 40% is what makes a word read as part of the photograph. |
| Blend | How the letters mix with the photo underneath: multiply, screen, overlay, soft light, difference, luminosity. | Screen for a light word on a dark photo, multiply for a dark word on a light one. Both keep the texture of the picture in the letters. |
| Outline | A stroke around each letter, with its own colour. The width is a percentage of the type size. | The photo behind the word is busy, or the word sits half on the sky and half on the ground. |
| Shadow | A soft halo, with its own colour. No offset, so it reads as separation rather than as a drop shadow from 2009. | The word and the background are close in tone. A dark halo behind a white word is the single most reliable fix on this page. |
None of the four browser editors we could get a photo into has either of these. Between them they offer font, size, colour, spacing, rotation and position, and nothing at all for the moment a white word lands on a bright sky. If your text has gone muddy over a photo, try Shadow at about 12 before you try a different colour.
Place: depth, rotation and the middle
Depth is the switch this whole tool is named after. Behind puts the layer under the cutout, in front puts it over the top. It is set per layer, so one picture can have a headline behind the subject and a small caption over their shoulder.
| Control | What it does |
|---|---|
| Depth | Behind or in front of the subject, for this layer only. |
| Rotate | Turns the layer about its own centre. The handle on the picture does the same thing by hand. |
| Centre across | Puts the layer on the vertical middle of the picture, leaving its height alone. |
| Centre down | Puts it on the horizontal middle, leaving its position across alone. |
Cutout: fixing the edge without re-running anything
The model runs once per photo, and its result is held. That is why these two sliders come back instantly: they reshape the matte that is already there rather than asking for a new one.
| Control | What it does | Use it when |
|---|---|---|
| Subject | Which model finds the shape. People is 13 MB, Anything else is 84 MB. | Switch to Anything else for an animal, a product, a car or an illustration. Changing it re-runs the model. |
| Edge | Moves the boundary out or in, a pixel at a time. | Negative when a rim of the old background came along. Positive when a shoulder or an ear got shaved off. |
| Soften | Blurs the boundary. | Hair, fur and anything else with a fuzzy edge. Two or three pixels is usually enough. |
Saving, and what the file is
The picture on screen and the file you save are drawn by the same function. The preview is that drawing at whatever size fits the panel; the export is the same drawing at the photograph’s own resolution. There is no second renderer to fall out of step, which is why the export can be promised to match what you were looking at.
| Format | Use it for |
|---|---|
| PNG | The default. Lossless, so the letters have no compression fuzz around them. |
| JPG | A smaller file for a photo-heavy design. Quality 0.92, and hard edges pick up a little ringing. |
| WebP | Smaller again at the same quality. The option is removed in Safari, which displays WebP and cannot write one. |
Keyboard shortcuts
| Keys | Does |
|---|---|
| Ctrl V or Cmd V | Open an image from the clipboard. |
| Arrow keys | Nudge the selected layer one pixel. Click the picture first. |
| Shift and an arrow | Nudge ten pixels. |
| Shift while rotating | Step the angle in fifteens. |
| Delete or Backspace | Remove the selected layer, with the picture focused. |
| Esc | Leave full screen. |
What this tool will not do
The model runs on your machine, which is the reason the photo can stay put and also the reason a few things are out of reach. Both halves are true.
- Video. One still image at a time. Text behind a moving object is a tracking job, and the guide covers where to do it.
- Two subjects, one behind and one in front. The cutout is a single layer, so everything the model found comes forward together.
- Painting the mask by hand. Edge and Soften reshape what the model found. There is no brush to add a region it missed.
- A batch. One picture at a time. The design is saved and can be restored onto the next photo, but they do not queue.
- Curved or warped text. Straight lines only, rotated as a block.
- Keeping your font. A font file lasts as long as the tab, because nothing about it is written to disk.
- Behaving well on a phone. See below. The cutout is the part that struggles, and it is not a small part.
Why this one wants a desktop
The rest of this site works as well on a phone as anywhere else, because the rest of this site is doing arithmetic. This tool is running a neural network, and that is a different kind of ask.
A phone gives a browser tab a small slice of memory and no thermal headroom. The people model at 12,984,781 bytes is usually survivable and slow. The general model at 88,117,930 bytes will run some phones out of memory rather than being merely slow, and what you get then is a reloaded tab rather than a message. Use Chrome on a desktop or a laptop for anything you care about.
Everything else here is fine on a phone. Dropping the picture, the type, the fonts, the blend modes and the export are ordinary canvas work, and the panel is built to be used with a thumb: the picture stays pinned to the top of the screen while the controls scroll under it, and the handles have a target twice the size of the dot you can see.
Where this beats the editors ranking above it
Seven editors for this effect were run on the same 1200x800 photograph on 4 September 2026, on an M1 Max in Chrome 152. The row where floi loses is in the table too.
| Editor | Sign-in | Ads | Size cap | Outline or shadow | Behind and in front |
|---|---|---|---|---|---|
| floi | No | None | None | Yes | Yes |
| textbehindimage.com | No | None | None | No | No |
| text-behind-image.org | No | None | 10 MB | No | No |
| textbehindimage.rexanwong.xyz | Google account | Sponsor banner | Not reached | Not reached | Not reached |
| underlayx.com | Google account | None | Not reached | Not reached | Not reached |
| textbehindimageai.com | No | None | None | No | Yes |
| canva.com | Canva account | None | Uploaded | Yes | Yes |
The row floi loses is Canva’s. It has a full type system, hundreds of licensed faces and a library of layouts, and this page has nine font stacks and a button that reads a file off your disk. What it also has is your photograph, on its servers, behind an account.
What "nothing is uploaded" means here, and how to check
It means the model that finds the subject is a file your browser downloads and then runs, the same way it runs the rest of this page. The photo goes into a function, the cutout comes out, and no request happens in between.
Two of the seven editors we opened for this page ask for a Google account before they will take a photo at all, and one of those discards the picture you had already chosen if you decline. The editor ranking first fetches 229 Google Fonts stylesheets on load, 11.02 MB of the 12.22 MB it pulls across 250 requests, before you have done anything.
Open the network panel, drop a picture in, and look. Here you will see the model load once and nothing after it. That test takes thirty seconds and it works on any tool that makes the claim.
Frequently asked questions
Is my photo uploaded anywhere?
No, and you can check rather than take our word for it. Open your browser's network panel and drop a picture in: the only request is for the cutout model, once, and after that a drop makes no request at all. Two of the seven editors we ran for this page want a Google account before they will open a photo, and a third fires a full-screen advert on the way in.
Why does the first photo take longer than the rest?
Because the model that finds the subject has to arrive before it can run. The people model is 12,984,781 bytes and downloads once, in 2.57 seconds on our connection, after which it is cached and the tool works offline. With it on the machine, the first cutout in a fresh tab took 2.16 seconds and every one after it took 1.53. The editor ranking first for this phrase took 71.96 seconds on its first photo and 38.20 on its second, measured the same day on the same machine with the same picture.
Which one should I pick under Subject?
People unless the thing in front is not a person. It is a human matting model, so on a leopard it is not slow, it is blank: it looks for a person, finds none, and returns nothing at all. Anything else handles animals, products, cars and illustrations. On our five test photos People was right on the four with a person in them and found nothing on the leopard, and Anything else was right on all five, taking 8.65 seconds a photo against 1.53.
Can I use my own font?
Yes, and none of the four browser editors we ran for this page offers it. Press Use a font file from this device, or drop a .ttf, .otf, .woff or .woff2 straight onto the panel, and it joins the font list. The file is read into memory with the browser's own FontFace and is never sent anywhere. It lasts as long as the tab does, so a design you come back to tomorrow needs the font adding again.
Does it put a watermark on the image?
No, and there is no export limit either. The file you save is the picture you dropped in, at its own resolution, with your text drawn into it. A 4032x3024 phone photo comes back at 4032x3024. The editor ranking second for this phrase stops accepting a file at 10 MB before you get that far.
The cutout ate part of the subject. Can I fix it?
Usually, with the two sliders under Cutout. Edge moves the boundary out or in a pixel at a time, which is the fix when a thin rim of background came along or a shoulder got shaved. Soften blurs the boundary, which is the fix for hair. Neither re-runs the model, so both are instant. If the shape is badly wrong rather than slightly wrong, try the other model first.
Can I put the text in front instead?
Yes. Every layer has a Depth switch, so some words can sit behind the subject and others in front of it in the same picture. That is the pair that reads as depth: a heading behind the head and a small caption over the shoulder. Of the four browser editors we could get a photo into, only one offers the choice at all.
Does it work on a phone?
It runs, and we would rather you did not rely on it. The cutout is a neural network executing inside the browser tab, and a phone gives that tab a small fraction of the memory a laptop does, with none of the cooling. The people model at 13 MB is usually survivable and slow. The 84 MB model runs some phones out of memory outright, and what you get is a reloaded tab rather than an error. Everything else on the page, the text, the fonts, the export, is ordinary canvas work and is fine. If the picture matters, open this in Chrome on a desktop or a laptop.
What happens to my design when I close the tab?
The recipe stays in this browser, and the photo does not. Saving an image writes the last ten designs to localStorage: the words, the face, the tracking, the colour and where each layer sits. No picture is ever stored, here or anywhere else. Because every measurement in a design is a fraction of the picture rather than a pixel count, restoring one onto a completely different photo puts the text in the right place.