diff --git a/src/h264.c b/src/h264.c
index 6efacb2..ff1585a 100644
--- a/src/h264.c
+++ b/src/h264.c
@@ -226,8 +226,12 @@ static inline int32_t h264_strip_ffmpeg_poc_sentinel(int32_t poc, uint32_t flags
 
 static void h264_fill_dpb(struct request_data *data,
 			  struct object_context *context,
+			  VAPictureParameterBufferH264 *VAPicture,
 			  struct v4l2_ctrl_h264_decode_params *decode)
 {
+	const int max_frame_num =
+		1 << (VAPicture->seq_fields.bits.log2_max_frame_num_minus4 + 4);
+	const int cur_frame_num = (int)VAPicture->frame_num;
 	int i;
 
 	for (i = 0; i < H264_DPB_SIZE; i++) {
@@ -246,7 +250,41 @@ static void h264_fill_dpb(struct request_data *data,
 		}
 
 		dpb->frame_num = entry->pic.frame_idx;
-		dpb->pic_num = entry->pic.picture_id;
+
+		/*
+		 * Per ext-ctrls-codec-stateless.rst, dpb[].pic_num must
+		 * equal the H.264 spec's PicNum (8-28) for short-term refs
+		 * or LongTermPicNum (8-29) for long-term refs.
+		 *
+		 * For frames (not field-coded), PicNum = FrameNumWrap.
+		 * FrameNumWrap = (frame_num > cur_frame_num)
+		 *                ? frame_num - max_frame_num
+		 *                : frame_num
+		 * (per spec section 8.2.4.1, frame_num wraparound).
+		 *
+		 * VAAPI convention (libavcodec/vaapi_h264.c::fill_vaapi_pic
+		 * line 64): VAPictureH264.frame_idx holds long_term_frame_idx
+		 * for long-term refs and frame_num for short-term refs. So
+		 * for long-term entries we copy frame_idx straight through
+		 * as LongTermPicNum.
+		 *
+		 * fourier's previous code set pic_num to picture_id (the
+		 * VAAPI surface id) which is unrelated to H.264 PicNum;
+		 * mediatek's vdec_h264_req_common.c::dst_entry->pic_num is
+		 * one consumer that fails on that. Hantro doesn't read
+		 * pic_num at all (uses reference_ts for ref resolution),
+		 * which is why fourier's wrong value never surfaced on
+		 * RK3568.
+		 */
+		if (entry->pic.flags & VA_PICTURE_H264_LONG_TERM_REFERENCE) {
+			dpb->pic_num = entry->pic.frame_idx;
+		} else {
+			int frame_num = (int)entry->pic.frame_idx;
+			dpb->pic_num = (frame_num > cur_frame_num)
+				? frame_num - max_frame_num
+				: frame_num;
+		}
+
 		dpb->top_field_order_cnt =
 			h264_strip_ffmpeg_poc_sentinel(entry->pic.TopFieldOrderCnt,
 						       entry->pic.flags);
@@ -283,7 +321,7 @@ static void h264_va_picture_to_v4l2(struct request_data *driver_data,
 	nal_ref_idc = (b[0] >> 5) & 0x3;
 	nal_unit_type = b[0] & 0x1f;
 
-	h264_fill_dpb(driver_data, context, decode);
+	h264_fill_dpb(driver_data, context, VAPicture, decode);
 
 	/*
 	 * Populate every V4L2_CID_STATELESS_H264_DECODE_PARAMS field