The line as the recogniser gets it, cut into columns of 8 pixels. Purple arcs: the columns the current one looks at (attention, second transformer block). Bars above: how sure each column is of its best symbol (grey: the blank, “no new character here”; red unsure → green sure). Below: each column's best symbol, with the runs CTC merges boxed. Hover the line to move the cursor.
Every stage's real output, stretched to the same width as the line, so the cursor marks the same place in every row: follow a letter from pixels to strokes to columns to attention to the symbol. Orange positive, blue negative.
The diagram lights up block by block while this panel shows what comes out of each stage.
A PDF is read whole, up to 20 pages. The layout model runs in your browser (130 MB the first time); each region is then read by PaddleOCR-VL-1.6 on a RunPod GPU. The first page after a pause waits about a minute while the GPU starts.
The diagram lights up block by block while this panel shows what comes out of each stage.