Code Image Decoder

1. Upload a picture of the coded message, or click Example to try one.
2. The decoder tries to identify the code for you on load. If it is unsure, it shows a row of Possible codes with a small picture of each. Click the one that looks like your image.
3. You can also set the code yourself in the Code family box: open it to browse the full list, or start typing and the matches narrow as you go. You do not need to know the exact name.
4. Click Read glyphs. The letters appear under each symbol and in the editable box below.
5. Fix anything that came out wrong, then send the text to another tool with the buttons under the result.

How the reading works: the image is converted to black and white and split into individual glyphs. Spacing is worked out automatically, so codes made of separate dots (like braille) are not broken apart at their internal gaps. Each glyph is trimmed, scaled to a common size, and compared against the reference symbols for the code you picked. The closest letter wins, and the runner-up is kept in case you need to switch.

Correcting a reading. Uncertain glyphs are outlined in amber. Click any glyph in the grid to open a picker that shows your symbol next to the code's reference symbols, then click the right one. If the split itself is wrong, use Edit boxes to drag a new box around a symbol or click a box to remove it. The reading updates as you go, and the text box is fully editable.

Tricky layouts. For a message that runs over several lines, set the number of rows in the Layout boxes. For gapless tiled codes that have no spacing between symbols (such as Maze), set both rows and columns after counting them on the enlarged image.

When some glyphs come back uncertain, Refine with model sends just those glyphs to the same image model that powers Identify the code. It re-ranks the candidate letters within the family using a richer comparison, which often resolves shapes that look alike to the quick local match. This step needs a connection; everything else runs in your browser.

Best results come from a straight-on, high-contrast image of evenly spaced symbols on a plain background. Hand-drawn or photographed codes with rotation, shadows, or uneven spacing are harder and may need a little manual correction. Colour-based codes (such as Hexahue and signal flags) are supported too. The decoded text is checked two ways as a sanity check: how much of it forms dictionary words, and whether it looks like a coordinate.