diff --git a/src/surface.c b/src/surface.c
index 0b8f802..31cba3c 100644
--- a/src/surface.c
+++ b/src/surface.c
@@ -118,10 +118,20 @@ VAStatus RequestCreateSurfaces2(VADriverContextP context, unsigned int format,
 
 		capture_type = v4l2_type_video_capture(video_format->v4l2_mplane);
 
-		rc = v4l2_set_format(driver_data->video_fd, capture_type,
-				     video_format->v4l2_format, width, height);
-		if (rc < 0)
-			return VA_STATUS_ERROR_OPERATION_FAILED;
+		/*
+		 * Do not VIDIOC_S_FMT on the CAPTURE queue. The hantro
+		 * stateless decoder derives the CAPTURE format from the
+		 * SPS attached to the OUTPUT request; explicitly setting
+		 * it here can put the driver into an inconsistent state.
+		 * GStreamer's v4l2slh264dec only G_FMTs CAPTURE (see
+		 * gst-plugins-bad/sys/v4l2codecs/gstv4l2decoder.c::
+		 * gst_v4l2_decoder_negotiate_src_format), and that
+		 * variant produces correct decoded NV12 on the same
+		 * hardware where this driver currently emits zeros.
+		 *
+		 * v4l2_get_format() below queries the driver's current
+		 * state and gives us the bytesperline/sizes we need.
+		 */
         } else {
 		video_format = driver_data->video_format;
 		capture_type = v4l2_type_video_capture(video_format->v4l2_mplane);