diff --git a/admin-lehrer/components/ocr-overlay/usePixelWordPositions.ts b/admin-lehrer/components/ocr-overlay/usePixelWordPositions.ts
index 1141181..b7b1b27 100644
--- a/admin-lehrer/components/ocr-overlay/usePixelWordPositions.ts
+++ b/admin-lehrer/components/ocr-overlay/usePixelWordPositions.ts
@@ -126,10 +126,20 @@ export function usePixelWordPositions(
 
         const wordPos: WordPosition[] = []
 
-        if (groups.length > 1 && clusters.length >= groups.length) {
-          // Multiple text groups with enough clusters: map each group to a cluster
+        // When more clusters than groups, pick the N widest clusters
+        // (text clusters are wider than box-border clusters)
+        let matchClusters = clusters
+        if (groups.length > 1 && clusters.length > groups.length) {
+          matchClusters = [...clusters]
+            .sort((a, b) => (b.end - b.start) - (a.end - a.start))
+            .slice(0, groups.length)
+            .sort((a, b) => a.start - b.start)
+        }
+
+        if (groups.length > 1 && matchClusters.length >= groups.length) {
+          // Multiple text groups with matching clusters
           for (let i = 0; i < groups.length; i++) {
-            const cl = clusters[i]
+            const cl = matchClusters[i]
             const clusterW = cl.end - cl.start + 1
             const measured = ctx.measureText(groups[i])
             const autoFontPx = refFontSize * (clusterW / measured.width)