diff --git a/src/surface.c b/src/surface.c
index 432a46a..58ca9cb 100644
--- a/src/surface.c
+++ b/src/surface.c
@@ -672,18 +672,64 @@ VAStatus RequestExportSurfaceHandle(VADriverContextP context,
 						      surface_object->destination_sizes[i];
 	}
 
-	surface_descriptor->num_layers = 1;
+	/*
+	 * Layer construction depends on the consumer's request flags
+	 * (VA_EXPORT_SURFACE_*_LAYERS):
+	 *
+	 *   COMPOSED_LAYERS (default, mpv): one layer carrying both
+	 *   Y and UV planes (drm_format=NV12, num_planes=2). Mesa
+	 *   imports as a single NV12 EGLImage.
+	 *
+	 *   SEPARATE_LAYERS (Firefox 150 RDD): two layers, Y as a
+	 *   single-plane R8 layer, UV as a single-plane GR88 layer.
+	 *   Firefox's GetVAAPISurfaceDescriptor passes
+	 *   VA_EXPORT_SURFACE_SEPARATE_LAYERS so its DMABufSurfaceYUV
+	 *   import code can address Y and UV planes independently.
+	 *   Without this branch, Firefox parsed our COMPOSED layout
+	 *   as if it were SEPARATE, found bogus layer-1 data, and
+	 *   silently fell back to FFmpeg(FFVPX) software decode.
+	 *
+	 * The earlier path 0001 mplane port assumed a single COMPOSED
+	 * shape — fine for mpv but breaks any consumer requesting
+	 * SEPARATE. Honor the flag.
+	 */
+	if ((flags & VA_EXPORT_SURFACE_SEPARATE_LAYERS) && planes_count == 2) {
+		surface_descriptor->num_layers = 2;
 
-	surface_descriptor->layers[0].drm_format = video_format->drm_format;
-	surface_descriptor->layers[0].num_planes = planes_count;
+		/* Layer 0: Y plane as DRM_FORMAT_R8 (1 byte/pixel luma). */
+		surface_descriptor->layers[0].drm_format = DRM_FORMAT_R8;
+		surface_descriptor->layers[0].num_planes = 1;
+		surface_descriptor->layers[0].object_index[0] =
+			export_fds_count == 1 ? 0 : 0;
+		surface_descriptor->layers[0].offset[0] =
+			surface_object->destination_offsets[0];
+		surface_descriptor->layers[0].pitch[0] =
+			surface_object->destination_bytesperlines[0];
 
-	for (i = 0; i < planes_count; i++) {
-		surface_descriptor->layers[0].object_index[i] =
-			export_fds_count == 1 ? 0 : i;
-		surface_descriptor->layers[0].offset[i] =
-			surface_object->destination_offsets[i];
-		surface_descriptor->layers[0].pitch[i] =
-			surface_object->destination_bytesperlines[i];
+		/* Layer 1: UV plane as DRM_FORMAT_GR88 (interleaved
+		 * U+V, 2 bytes/pixel chroma at half resolution). */
+		surface_descriptor->layers[1].drm_format = DRM_FORMAT_GR88;
+		surface_descriptor->layers[1].num_planes = 1;
+		surface_descriptor->layers[1].object_index[0] =
+			export_fds_count == 1 ? 0 : 1;
+		surface_descriptor->layers[1].offset[0] =
+			surface_object->destination_offsets[1];
+		surface_descriptor->layers[1].pitch[0] =
+			surface_object->destination_bytesperlines[1];
+	} else {
+		/* COMPOSED_LAYERS / default: one layer with all planes. */
+		surface_descriptor->num_layers = 1;
+		surface_descriptor->layers[0].drm_format = video_format->drm_format;
+		surface_descriptor->layers[0].num_planes = planes_count;
+
+		for (i = 0; i < planes_count; i++) {
+			surface_descriptor->layers[0].object_index[i] =
+				export_fds_count == 1 ? 0 : i;
+			surface_descriptor->layers[0].offset[i] =
+				surface_object->destination_offsets[i];
+			surface_descriptor->layers[0].pitch[i] =
+				surface_object->destination_bytesperlines[i];
+		}
 	}
 
 	/*