Every file you upload goes through our image to PPT OCR recogniser on our own servers before it becomes a slide. The more useful question is whether it will succeed on your page, and that is decided by things you can check before you upload.
Choose up to 20 JPG, PNG, WebP, or SVG files from your Mac, PC, iPad, or phone
Three parts, and the middle one decides how your page turns out.
The engine is Tesseract, running in its neural recognition mode on our own servers. Nothing is sent to an outside provider, and in this free release no generative model is involved at any point.
Recognition returns a candidate for every line along with how sure it is. Lines below the bar are discarded rather than written into the deck, and where a line is discarded its original pixels are left in place.
Holding a stack of scans and wondering whether this is worth doing?
Pick the least promising page you have and run that one alone. Nothing written here beats seeing what your own material does.
The reading switch is on by default. Turning it off skips recognition entirely and places each file as a plain picture, which takes less time when you have no text to edit.
Check your hardest scan first before queuing the rest of the folder.
Recognition is enabled by default; turn it off only when you want flat picture slides without text extraction.
Run your hardest page on its own first, or upload up to 20 files at 20 MB each in one batch.
Read the completion panel for any unreadable-page flags, then open the deck and click into a heading.
None of these is a threshold you can look up, because the engine weighs them together.
Pixel height of the lettering, and contrast against the background. Characters need enough pixels to carry their shape, and body copy you already struggle to read at full size will come back as a picture; dark grey on mid grey, a watermark behind a paragraph or a photograph showing through all reduce the separation the engine needs.
Straightness, and whether the writing is printed. A scan sitting a few degrees off curves the baselines, and the engine reads in lines; marker pen, biro and signatures are not what a recogniser trained on typeset characters learned.
And which script it is in. The installed packs cover English and Simplified Chinese, which between them supply the Latin and Han character shapes. Scripts built from neither, Japanese kana, Korean hangul, Cyrillic and Arabic among them, normally stay as picture.
Converting slide images to PowerPoint lets you edit headings, update metrics, and reuse existing slide layouts without retyping every text box from scratch.
Drops one flattened bitmap onto a blank slide. You still have to retype every title, bullet, and number manually.
Rebuilds native PowerPoint text frames at exact slide coordinates while keeping photographic plates crisp.
Where text is too small, curved, or low-contrast to read safely, we keep that region as your original image instead of guessing gibberish characters.
Every reconstructed text box is anchored to the exact X/Y coordinates of your original slide so multi-column layouts and side-by-side comparisons stay aligned.
Your images are processed in volatile RAM and discarded as soon as your PowerPoint is delivered, or within 10 minutes if not downloaded.
All five factors are visible to you before you upload.
Large, high-contrast lettering a computer rendered for a display is the best case this engine has.
A page fed in square, in focus and evenly lit gives the reader parallel baselines to work along.
Characters need enough pixels to carry their shape, and enlarging the file afterwards adds no detail that was not captured.
Pale type on a tinted panel, or a watermark behind a paragraph, reduces the separation the engine depends on.
A photograph taken at an angle bends the baselines, and a curved line pushes more of the page below the bar.
Latin and Han character sets are supported; lines written in kana, hangul, Cyrillic or Arabic stay inside the slide image.
Anyone holding material that may or may not survive the reading step.
Two hundred scanned pages, and one representative page answers whether the set is worth running.
Figures and captions from papers, where a misread caption would be worse than a picture of one.
Old handouts that exist only on paper, with the headings worth recovering as text.
Documents where an invented word is unacceptable, so a silent guess is the wrong failure.
Interface captures, which are the case that reads back most completely.
Check whether scanned headings extract cleanly before committing to a full translation workflow.
Coming up short is a normal outcome rather than an error, so it is worth knowing what it looks like.
A page that is one flat colour is detected as empty before recognition starts, so no time is spent reading it and nothing is invented to fill it.
When no line on a page clears the bar, the result panel marks that file with a note saying no readable text was found and it was kept as an image. You see this before opening the deck.
The case to watch is the partial one: a page where most lines rebuilt but a watermark or a caption did not gets no marker, because the report only separates all from nothing. Open those pages and check the small print yourself.
Go straight to the converter page built for your source format.
Yes. Unless you turn the recognition toggle off before starting, every non-blank upload is scanned for printed text before the slide package is assembled.
We run Tesseract in neural LSTM mode directly on our own infrastructure. No external cloud vision API or generative language model sees your files.
Server recognition ships with English (`eng`) and Simplified Chinese (`chi_sim`) training data. That pairing reads Latin alphabets and Simplified Han characters, while Japanese kana, Korean hangul, Cyrillic and Arabic remain part of the slide picture.
Recognition depends on five combined properties of the source file: character pixel height, foreground-to-background contrast, horizontal baseline alignment, printed rather than handwritten strokes, and script coverage.
What matters is the pixel height of the smallest letters rather than the overall canvas dimensions. If fine print looks blurry when viewed at actual size on your monitor, upscaling the image file beforehand will not recover the missing strokes.
Low-confidence lines are not guessed or written as garbled text boxes. Instead, those regions remain visible as part of the original slide picture so nothing fabricated slips past review.
The download panel lists every processed file and attaches a notice to any slide where zero lines cleared the confidence threshold. Slides where headings converted but small footnotes stayed pictorial do not show a warning badge, so inspect fine print after opening the deck.
Scans sent through the recogniser live in memory only while Tesseract scores the lines, and are wiped when you collect the .pptx or when the 10-minute window closes. No page data is retained.
Run one representative page and read the panel. That answers the question this page is about better than any claim on it.
Try One Page
