diff --git a/src/context.c b/src/context.c
index 2c863c8..d959d1b 100644
--- a/src/context.c
+++ b/src/context.c
@@ -146,6 +146,23 @@ VAStatus RequestCreateContext(VADriverContextP context, VAConfigID config_id,
 					dev_ctrls, 2);
 	}
 
+	/*
+	 * Mirror the ANNEX_B start-code mode set on the device above
+	 * into context_object->h264_start_code so picture.c::
+	 * codec_store_buffer prepends 0x00 0x00 0x01 to each slice
+	 * payload it copies into the OUTPUT buffer. Without this, the
+	 * kernel — which we just told to expect ANNEX_B — sees a raw
+	 * NAL stream with no start codes, fails to find slice
+	 * boundaries, and emits a zeroed CAPTURE buffer (visually a
+	 * flat dark-green frame).
+	 *
+	 * h264_get_controls() exists for this purpose but is never
+	 * called in the current code path; the planned probe-then-set
+	 * commit will replace this hardcoded assignment with a runtime
+	 * read of the kernel's accepted START_CODE value.
+	 */
+	context_object->h264_start_code = true;
+
 	rc = v4l2_set_stream(driver_data->video_fd, output_type, true);
 	if (rc < 0) {
 		status = VA_STATUS_ERROR_OPERATION_FAILED;