Nine jobs for your photos, and why none of them need a server
This guide walks through everything the Photo AI Studio can do to a picture. Four jobs edit: cut the background away, put a colour behind the subject, blur the distance like a portrait lens, and make a picture bigger. Five jobs describe: write a title and alt text, read the words inside an image, count the objects in it, suggest keywords, and find the shots that are copies of each other. All of it runs on your own machine.
Two kinds of job on one page
The row of buttons across the top picks the job, and the jobs come in two families. The editing family - Remove background, Replace background, Portrait blur, and Enlarge - takes one picture at a time, because each of these is heavy work and the result is a new image you download. The describing family - Describe, Read text, Find objects, Keywords, and Find duplicates - takes as many pictures as you like and answers with words rather than pixels.
The page swaps its input area to match. Pick an editing job and you get a single-picture drop box that accepts JPG, PNG, WebP and GIF. Pick a describing job and the drop box invites a whole batch, shows a thumbnail of every picture you add, and works through them in order when you press Run.
The short line under the buttons always says what the chosen job will do, so you never have to remember which name means what.
The first use of a job is the slow one
Each job leans on its own piece of AI, and none of it is fetched until something needs it. The moment you pick a job, the page quietly starts collecting that job's piece in the background, so by the time your picture is loaded there is usually nothing left to wait for.
The pieces are not small - the describing jobs in particular lean on a large one - so the very first use of a family can take a minute on a slow connection. It happens once. After that the piece lives in your browser, later visits start at once, and every job keeps working with the internet switched off.
A small panel under the tool lists what has been stored this way and has a button to clear it out if you want the space back.
Removing and replacing a background
Remove background decides which pixels belong to the subject and drops everything else. The empty area is shown as grey and white squares in the preview; those squares only mark where the picture is now see-through, and the downloaded PNG is genuinely transparent there, so the cut-out sits cleanly on any colour or over any other picture.
The edges get extra care after the cut. The mask is tightened by about a pixel and softened, and the colour of the old background is mathematically subtracted from the edge pixels, so the pale rim that background removers usually leave around a subject is cleaned away on its own.
Replace background does the same cut and then fills the empty area with a flat colour. Its one control is a colour picker that starts on white. It is the fast route to a catalogue shot or an ID-style photo on a plain backdrop.
Portrait blur follows distance, not a formula
A flat blur smears the whole frame evenly and the eye notices at once. This job first works out how far away each part of the picture is, then keeps the near things sharp and softens what sits further back - which is how a real lens behaves.
Its control is a slider from 2 to 24, starting at 8. Run the job first, then drag: the strength changes the instant you move it, with no second run. Low numbers give a gentle falloff, high numbers melt the background away.
Enlarge has two speeds, and is never smaller
Pictures up to 1280 pixels on the longest side are redrawn by the AI at double size. The picture is worked through in parts, a few seconds each, and the bar counts them off as it goes, with an honest time estimate shown before you commit.
Bigger pictures skip the AI and are doubled instantly with a high-quality resample and a light sharpen, as large as the canvas on your device allows - a desktop usually manages inputs around 5000 pixels, a phone less. Whichever path runs, the result is always exactly twice the size of what you gave it, never smaller.
If a picture is too big even for the fast path, the note under the job says so before you run anything, and tells you the ceiling on the device you are using.
Describe writes three boxes, and they differ on purpose
Describe fills three editable boxes for each picture. The Title is short and headline-shaped. The Alt text is one plain sentence, which is what a screen reader should speak and what search engines index. The Description is longer and covers more of the frame.
Keep the first two apart. A title labels a picture; alt text describes it to someone who cannot see it. Swapping them is the most common accessibility mistake on the web, and having both written for you side by side is the easiest way to stop making it.
Every box accepts typing. Read what the AI wrote, fix a word, and the corrected version is what you copy out or export.
Reading text, counting objects, ranking keywords
Read text pulls any writing out of a picture into a single box - a sign, a label, a screenshot, a few words on a poster. For dense typed pages, receipts and forms, the separate OCR tool is the better fit; this job is tuned for words in the wild. If there is nothing to read it says so plainly.
Find objects draws a box around each thing it recognises, lays the boxed-up picture in the preview, and lists a tally underneath - person three times, car twice. It reports what it sees and can be confidently wrong about unusual pictures, so treat the count as a first pass.
Keywords ranks stock-photo tags for the picture. Each tag is a button that copies itself when clicked, a text box holds all of them comma-separated, and a note names the best-fitting category.
Finding duplicates, and the alike slider
Find duplicates needs at least two pictures, and its Run button says Compare pictures to match. It turns every photo into a long list of numbers that stands for how the photo looks, compares the lists, and groups the shots whose lists are close. Each group shows how alike its members are as a percentage.
One slider belongs to this job alone: count as duplicates above a chosen percentage, from 70 to 99, starting at 92. Lower it to catch cropped or edited copies of one shot; raise it to group only near-identical frames. If nothing matches, the tool says every picture looked different - a real answer, not a fault.
This is also why there is no box for searching photos by typed words. Matching words to pictures needs a piece of AI this page does not carry; matching pictures to pictures needs only what it does.
What comes out at the end
The editing jobs end in a Download PNG button, with the preview and the saved file drawn from the same canvas so they can never disagree. PNG is used because it can hold real transparency, which the background jobs depend on.
The describing jobs end in editable text. With one picture you copy from the boxes; with a batch, a Download CSV button appears and saves one row per picture, with your edits included. The duplicate job has no CSV, because its answer is groups of pictures rather than a table.
Clear empties the input so you can start over. Nothing you did is kept anywhere once the tab closes.
A worked example: thirty product shots
Say you photographed a batch of ceramic mugs for a shop. Start with Find duplicates: drop all thirty photos in, press Compare pictures, and two groups appear - the shots you accidentally took twice. Drop the weaker frame of each pair.
Switch to Describe with the survivors still loaded. Run once, and the tool works down the batch, filling titles, alt text and descriptions while you watch. Fix the two that came out odd, press Download CSV, and the listing text for the whole batch is one file.
Finish with the hero image: pick Replace background, load the best mug shot, choose a light grey, and download a catalogue-ready PNG. Three jobs, one page, and no photo went anywhere.
Common mistakes
- Loading one picture into Find duplicates. It compares pictures with each other, so it needs at least two - the button reminds you.
- Reading the grey squares as part of the image. They only mark transparency; the saved PNG is clear there.
- Putting the title into the alt-text box. The full sentence belongs in alt text; the short label is the title.
- Publishing a description unread. The AI reports what it sees and is sometimes sure of itself while wrong.
- Expecting server speed. Everything runs on your own processor; seconds of waiting are the price of photos that never travel.
- Using Read text on a dense document. The OCR tool is built for pages of type; this job shines on a few words in a photo.
Why switching off the internet changes nothing
Every piece of AI this page uses is a file the site hands your browser once, on first use. From then on the work happens on the chip in front of you, and no part of any picture is ever sent out. Client photos, unpublished work, pictures of your family - none of it can leak, because there is no server in the loop to leak from.
You do not have to take that on trust. Load the page, use a job once so its piece is stored, then disconnect from the internet. Every job on the page keeps working exactly as before.