Study guides · Updated October 2026

Image to text: how to get a transcript you can actually trust

Converting an image to text sounds like a solved problem until the output arrives with the caption wedged into the second paragraph and a zero where an O should be. Recognising the letters is the easy part; deciding what order they go in, and knowing which characters to re-read afterwards, is where a transcript is won or lost. This guide is about that second half.

PNG, JPEG and WEBP. Free to try with the default model.

Ask for the text, or you will get the gist

The same upload behaves differently depending on what you type. An empty prompt box returns a summary, which is the wrong output if you came here for a quotation. Asking for all the readable text returns the OCR-style transcript instead: the words as they appeared, in reading order, with nothing paraphrased away. One sentence of instruction is the whole switch, and it is the single most common reason people think the extraction failed when it did not.

If you want both, take them in two passes on the same image. The comparison between the two outputs, and why they exist at all, is laid out in the image summary guide.

Reading order is the hard problem, not the letters

Modern recognition handles clean printed characters well. What no reader can know for certain is the sequence a human eye would follow. A page with two columns, a boxed definition, a figure caption and a running header has at least four candidate reading orders, and only you know which one matches the argument you are trying to follow.

Crop to one column and the ambiguity disappears, because there is only one way to read what is left. Two narrow uploads reliably beat one wide one on a newspaper page, a journal article, or anything laid out in a textbook with a coloured sidebar running down the edge. The same logic applies to tables: crop to the table alone and ask for it as a table, and the rows stay attached to their headings.

The characters to re-read before you paste it anywhere

Proofreading a long transcript word by word defeats the point of extracting it. Proofread selectively instead, because mistakes cluster in predictable places:

Prose protects itself: a misread letter in an ordinary sentence usually produces a word that is obviously wrong. Reference numbers, measurements and names have no such safety net, which is why they are worth the thirty seconds of checking.

Screenshot to text beats photographing a screen

If the words you want are already displayed on a device, capture them with a screenshot rather than pointing a second camera at the glass. A photo of a monitor adds reflections, a moire pattern from the pixel grid, and a keystone distortion from the angle you were sitting at, and all three are problems that simply do not exist in a screen capture.

The same preference holds for slides inside a lecture recording, where pausing and capturing gives a far cleaner source than a photo taken in the room. That workflow has its own guide in screenshot notes from video.

How to check a transcript you cannot read yourself

Sometimes the original is the problem — faded print, an unfamiliar script, handwriting that is not yours. You cannot proofread against a source you cannot decipher, so verify a different way: crop hard to the one element you doubt — the figure in the third row, the word before a heading — and run that crop as a new request on its own. A character read in isolation, with nothing around it to bias the guess, either agrees with the first attempt or it does not, and a disagreement is the signal to go back to the original.

Do this as a fresh upload rather than as a follow-up question. The follow-up box carries the conversation, not the picture, so a question asked there is answered from the text already on screen — which is the one thing you are trying to check.

Where the stakes are high — a dosage, a grade, a legal reference — treat the transcript as a draft to be confirmed against the original by someone who can read it, not as the record itself.

Where the text goes next

Extraction is rarely the end of the job. Once the words exist you can ask for them shortened into notes, which is the route covered in the image summarizer guide, or you can ask a question about what the page means, which is picture answer. All three start from the same upload on Image Summary.

Cropping happens in your browser, so only the region you kept is ever uploaded, and the request itself is stateless — the image is discarded once your answer is written rather than stored against your account. The privacy policy has the specifics.

Image to text converter FAQ