diff --git a/mpp/codec/dec/avs2/avs2d_dpb.c b/mpp/codec/dec/avs2/avs2d_dpb.c
index b2fbe406..e6ae58b2 100644
--- a/mpp/codec/dec/avs2/avs2d_dpb.c
+++ b/mpp/codec/dec/avs2/avs2d_dpb.c
@@ -323,9 +323,17 @@ static MPP_RET dpb_remove_frame(Avs2dCtx_t *p_dec, Avs2dFrame_t *p)
     MppFrame frame = p->frame;
 
     avs2d_dbg_dpb("In.");
+
+    if (p->picture_type == GB_PICTURE || p->invisible) {
+        MppBuffer buffer = NULL;
+        mpp_buf_slot_get_prop(p_dec->frame_slots, p->slot_idx, SLOT_BUFFER, &buffer);
+        mpp_buffer_put(buffer);
+    }
+
     mpp_buf_slot_clr_flag(p_dec->frame_slots, p->slot_idx, SLOT_CODEC_USE);
     avs2d_dbg_dpb("dpb remove frame slot_idx %d, doi %d poi %d, dpb used %d",
                   p->slot_idx, p->doi, p->poi, mgr->used_size);
+
     memset(p, 0, sizeof(Avs2dFrame_t));
     p->frame = frame;
     p->slot_idx = NO_VAL;
@@ -467,7 +475,7 @@ static Avs2dFrame_t *dpb_alloc_frame(Avs2dCtx_t *p_dec, HalDecTask *task)
     frm->intra_frame_flag = (frm->scene_frame_flag || frm->picture_type == I_PICTURE);
     frm->refered_by_scene = frm->scene_frame_flag;
     frm->refered_by_others = (frm->picture_type != GB_PICTURE && mgr->cur_rps.refered_by_others);
-    avs2d_dbg_dpb("frame picture type ref by others %d\n", frm->refered_by_others);
+    avs2d_dbg_dpb("frame picture type %d, ref by others %d\n", frm->picture_type, frm->refered_by_others);
     if (vsh->chroma_format == CHROMA_420 && vsh->bit_depth == 8) {
         mpp_frame_set_fmt(mframe, MPP_FMT_YUV420SP);
     } else if (vsh->chroma_format == CHROMA_420 && vsh->bit_depth == 10) {
@@ -677,10 +685,10 @@ static MPP_RET dpb_set_frame_refs(Avs2dCtx_t *p_dec, Avs2dFrameMgr_t *mgr, HalDe
         if (!mgr->scene_ref) {
             error_flag = 1;
         } else if (mgr->scene_ref != mgr->refs[0] || mgr->num_of_ref > 1) {
-            AVS2D_DBG(AVS2D_DBG_ERROR, "Error reference frame(doi %lld ~ %lld) for S.\n",
+            AVS2D_DBG(AVS2D_DBG_ERROR, "Error reference frame(doi %ld ~ %ld) for S.\n",
                       mgr->scene_ref->doi, mgr->refs[0] ? mgr->refs[0]->doi : -1);
-            mgr->num_of_ref = 1;
             replace_ref_flag = 1;
+            p_dec->syntax.refp.scene_ref_replace_pos = 0;
         }
     } else if ((p_cur->picture_type == P_PICTURE || p_cur->picture_type == F_PICTURE) &&
                p_dec->ph.background_reference_flag) {
@@ -689,6 +697,7 @@ static MPP_RET dpb_set_frame_refs(Avs2dCtx_t *p_dec, Avs2dFrameMgr_t *mgr, HalDe
             error_flag = 1;
         } else {
             replace_ref_flag = 1;
+            p_dec->syntax.refp.scene_ref_replace_pos = mgr->num_of_ref - 1;
         }
     } else if (p_cur->picture_type == B_PICTURE &&
                (mgr->num_of_ref != 2 || (mgr->refs[0] && mgr->refs[0]->poi <= p_cur->poi) ||
@@ -724,10 +733,8 @@ MPP_RET avs2d_dpb_insert(Avs2dCtx_t *p_dec, HalDecTask *task)
 
     AVS2D_PARSE_TRACE("In.");
 
-    if (p_dec->ph.picture_type != GB_PICTURE) {
-        //!< output frame from dpb
-        dpb_scan_output_frame(p_dec);
-    }
+    //!< output frame from dpb
+    dpb_scan_output_frame(p_dec);
 
     compute_frame_order_index(p_dec);
 
diff --git a/mpp/codec/dec/avs2/avs2d_ps.c b/mpp/codec/dec/avs2/avs2d_ps.c
index ec1788a6..8fc19321 100644
--- a/mpp/codec/dec/avs2/avs2d_ps.c
+++ b/mpp/codec/dec/avs2/avs2d_ps.c
@@ -591,6 +591,7 @@ static MPP_RET parse_picture_header_intra(Avs2dCtx_t *p_dec)
     BitReadCtx_t *bitctx  = &p_dec->bitctx;
     Avs2dSeqHeader_t *vsh = &p_dec->vsh;
     Avs2dPicHeader_t *ph  = &p_dec->ph;
+    RK_U32 last_pic_out_delay = ph->picture_output_delay;
 
     AVS2D_PARSE_TRACE("In.");
     memset(ph, 0, sizeof(Avs2dPicHeader_t));
@@ -627,7 +628,7 @@ static MPP_RET parse_picture_header_intra(Avs2dCtx_t *p_dec)
         AVS2D_PARSE_TRACE("temporal_id %d\n", ph->temporal_id);
     }
 
-    if (vsh->low_delay == 0 && (!ph->background_picture_flag || ph->background_picture_output_flag)) {
+    if (vsh->low_delay == 0 && !(ph->background_picture_flag && !ph->background_picture_output_flag)) {
         READ_UE(bitctx, &ph->picture_output_delay);
         AVS2D_PARSE_TRACE("picture_output_delay %d\n", ph->picture_output_delay);
         if (ph->picture_output_delay >= 64) {
@@ -636,7 +637,7 @@ static MPP_RET parse_picture_header_intra(Avs2dCtx_t *p_dec)
             goto __FAILED;
         }
     } else {
-        ph->picture_output_delay = 0;
+        ph->picture_output_delay = last_pic_out_delay;
     }
 
     READ_ONEBIT(bitctx, &predict);
@@ -737,6 +738,7 @@ static MPP_RET parse_picture_header_inter(Avs2dCtx_t *p_dec)
 
     if (vsh->low_delay == 0) {
         READ_UE(bitctx, &ph->picture_output_delay);
+        AVS2D_PARSE_TRACE("picture_output_delay %d\n", ph->picture_output_delay);
         if (ph->picture_output_delay >= 64) {
             ret = MPP_NOK;
             mpp_err_f("invalid picture output delay(%d) intra.\n", ph->picture_output_delay);
diff --git a/mpp/common/avs2d_syntax.h b/mpp/common/avs2d_syntax.h
index adb8fe5a..d44bb87c 100644
--- a/mpp/common/avs2d_syntax.h
+++ b/mpp/common/avs2d_syntax.h
@@ -57,6 +57,7 @@ typedef struct _DXVA_RefParams_AVS2 {
     RK_S32      ref_poc_list[32];
     RK_S32      scene_ref_slot_idx;
     RK_S32      scene_ref_enable;
+    RK_S32      scene_ref_replace_pos;
 } RefParams_Avs2d, *LP_RefParams_Avs2d;
 
 typedef struct _DXVA_AlfParams_AVS2 {
diff --git a/mpp/hal/rkdec/avs2d/hal_avs2d_rkv.c b/mpp/hal/rkdec/avs2d/hal_avs2d_rkv.c
index 3f4ce685..2c134540 100644
--- a/mpp/hal/rkdec/avs2d/hal_avs2d_rkv.c
+++ b/mpp/hal/rkdec/avs2d/hal_avs2d_rkv.c
@@ -427,14 +427,15 @@ static MPP_RET fill_registers(Avs2dHalCtx_t *p_hal, Vdpu34xAvs2dRegSet *p_regs,
         if (p_hal->syntax.refp.scene_ref_enable && p_hal->syntax.refp.scene_ref_slot_idx >= 0) {
             MppFrame scene_ref = NULL;
             RK_S32 slot_idx = p_hal->syntax.refp.scene_ref_slot_idx;
+            RK_S32 replace_idx = p_hal->syntax.refp.scene_ref_replace_pos;
 
             mpp_buf_slot_get_prop(p_hal->frame_slots, slot_idx, SLOT_FRAME_PTR, &scene_ref);
 
             if (scene_ref) {
-                p_regs->avs2d_addr.ref_base[refp->ref_pic_num - 1] = get_frame_fd(p_hal, slot_idx);
+                p_regs->avs2d_addr.ref_base[replace_idx] = get_frame_fd(p_hal, slot_idx);
                 mv_buf = hal_bufs_get_buf(p_hal->cmv_bufs, slot_idx);
-                p_regs->avs2d_addr.colmv_base[refp->ref_pic_num - 1] = mpp_buffer_get_fd(mv_buf->buf[0]);
-                p_regs->avs2d_param.reg67_098_ref_poc[refp->ref_pic_num - 1] = mpp_frame_get_poc(scene_ref);
+                p_regs->avs2d_addr.colmv_base[replace_idx] = mpp_buffer_get_fd(mv_buf->buf[0]);
+                p_regs->avs2d_param.reg67_098_ref_poc[replace_idx] = mpp_frame_get_poc(scene_ref);
             }
         }
 
diff --git a/mpp/hal/rkdec/avs2d/hal_avs2d_vdpu382.c b/mpp/hal/rkdec/avs2d/hal_avs2d_vdpu382.c
index abfa6b8d..8b0945a5 100644
--- a/mpp/hal/rkdec/avs2d/hal_avs2d_vdpu382.c
+++ b/mpp/hal/rkdec/avs2d/hal_avs2d_vdpu382.c
@@ -429,15 +429,16 @@ static MPP_RET fill_registers(Avs2dHalCtx_t *p_hal, Vdpu382Avs2dRegSet *p_regs,
 
         if (p_hal->syntax.refp.scene_ref_enable && p_hal->syntax.refp.scene_ref_slot_idx >= 0) {
             MppFrame scene_ref = NULL;
+            RK_S32 replace_idx = p_hal->syntax.refp.scene_ref_replace_pos;
             RK_S32 slot_idx = p_hal->syntax.refp.scene_ref_slot_idx;
 
             mpp_buf_slot_get_prop(p_hal->frame_slots, slot_idx, SLOT_FRAME_PTR, &scene_ref);
 
             if (scene_ref) {
-                p_regs->avs2d_addr.ref_base[refp->ref_pic_num - 1] = get_frame_fd(p_hal, slot_idx);
+                p_regs->avs2d_addr.ref_base[replace_idx] = get_frame_fd(p_hal, slot_idx);
                 mv_buf = hal_bufs_get_buf(p_hal->cmv_bufs, slot_idx);
-                p_regs->avs2d_addr.colmv_base[refp->ref_pic_num - 1] = mpp_buffer_get_fd(mv_buf->buf[0]);
-                p_regs->avs2d_param.reg67_098_ref_poc[refp->ref_pic_num - 1] = mpp_frame_get_poc(scene_ref);
+                p_regs->avs2d_addr.colmv_base[replace_idx] = mpp_buffer_get_fd(mv_buf->buf[0]);
+                p_regs->avs2d_param.reg67_098_ref_poc[replace_idx] = mpp_frame_get_poc(scene_ref);
             }
         }