diff --git a/CHANGELOG.md b/CHANGELOG.md
index 564ff81..430a5fb 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,7 @@
 
 - Decoder: 3-5x faster on small symbols and 1.5-2x on camera frames with identical results: word-wise luma conversion, pyramid and binarizer, finder runs walked straight off the packed row with exact early-outs, call-free grid sampling, packed codeword extraction, Reed-Solomon over packed words with syndromes taken from the remainder, and per-call arenas sized to the symbol instead of Version 40
 - Decoder: opt-in `nativeLimit` (and `nativeEvery` in `QRCanvas`) skips the full-resolution finder search on large camera frames
+- Decoder: read mirror images, by retrying a failed read on the transposed grid; iOS WebKit hands the rear-camera plane mirrored with no metadata saying so
 - Encoder: byte-identical output 10-30% faster for `raw`, `ascii`, `gif` and `data-url`, up to 3x for `svg`
 - DOM: fix the native VideoFrame path being disabled for a stream started before its first frame; do not overwrite the luma arena while an async decode reads it; mute through the property for mobile autoplay
 
diff --git a/src/decode.ts b/src/decode.ts
index 8e5b516..6d95108 100644
--- a/src/decode.ts
+++ b/src/decode.ts
@@ -2174,7 +2174,27 @@ export class _QRScanner {
   }
 
   // Parse format/function bits, deinterleave and correct codewords, then decode the payload.
-  private decodeGrid(size: number, _ctx: Ctx): Attempt {
+  // A mirror image locates like any symbol, since finders and timing are symmetric, and
+  // fails only at the format and data bits, so a failed read is retried on the transposed
+  // grid, which is the upright symbol. The symbology expects a reader to read mirror images;
+  // a camera plane can arrive mirrored with no metadata saying so.
+  private decodeGrid(size: number, ctx: Ctx): Attempt {
+    const result = this.decodeOriented(size, ctx);
+    if (!(result instanceof Error)) return result;
+    const grid = this.grid;
+    for (let y = 1; y < size; y++)
+      for (let x = 0; x < y; x++) {
+        const a = y * size + x;
+        const b = x * size + y;
+        const t = grid[a];
+        grid[a] = grid[b];
+        grid[b] = t;
+      }
+    const mirrored = this.decodeOriented(size, ctx);
+    return mirrored instanceof Error ? result : mirrored;
+  }
+
+  private decodeOriented(size: number, _ctx: Ctx): Attempt {
     let decoded: Attempt = FAIL.format;
     if (!checkVersion(this.grid, size)) decoded = FAIL.version;
     else {
diff --git a/test/decode.test.ts b/test/decode.test.ts
index cca6a78..9be49a4 100644
--- a/test/decode.test.ts
+++ b/test/decode.test.ts
@@ -1895,4 +1895,25 @@ it('decodeQR nativeLimit searches large frames on the half layer only', () => {
   deepStrictEqual(readQR(place(1, 100), { nativeLimit: 0 }), text);
 });
 
+it('decodeQR reads mirror images', () => {
+  // ISO/IEC 18004 expects a reader to read mirror images: a camera plane can arrive mirrored
+  // with no metadata saying so. Mirror a raster and a photo left to right.
+  const mirror = (img: { width: number; height: number; data: Uint8Array }) => {
+    const data = new Uint8Array(img.data.length);
+    const bpp = img.data.length / (img.width * img.height);
+    for (let y = 0; y < img.height; y++)
+      for (let x = 0; x < img.width; x++)
+        for (let c = 0; c < bpp; c++) {
+          const from = (y * img.width + img.width - 1 - x) * bpp + c;
+          data[(y * img.width + x) * bpp + c] = img.data[from];
+        }
+    return { width: img.width, height: img.height, data };
+  };
+  const text = 'MIRROR IMAGE 2026';
+  const raster = matrixToImage(encodeQR(text, 'raw', { border: 4 }), 4);
+  deepStrictEqual(readQR(mirror(raster)), text);
+  const photo = readImage('detection/blurred/image007.jpg');
+  deepStrictEqual(readQR(mirror(photo)), readQR(photo));
+});
+
 it.runWhen(import.meta.url);