diff --git a/PyNvCodec/TC/inc/NvDecoder.h b/PyNvCodec/TC/inc/NvDecoder.h
index e50dd511..2f0a74be 100644
--- a/PyNvCodec/TC/inc/NvDecoder.h
+++ b/PyNvCodec/TC/inc/NvDecoder.h
@@ -122,4 +122,5 @@ class DllExport NvDecoder {
   int ReconfigureDecoder(CUVIDEOFORMAT *pVideoFormat);
 
   struct NvDecoderImpl *p_impl;
+  bool setEndOfPicture = false;
 };
diff --git a/PyNvCodec/TC/src/NvDecoder.cpp b/PyNvCodec/TC/src/NvDecoder.cpp
index f8b18763..f01072d6 100644
--- a/PyNvCodec/TC/src/NvDecoder.cpp
+++ b/PyNvCodec/TC/src/NvDecoder.cpp
@@ -101,6 +101,11 @@ static int GetChromaPlaneCount(cudaVideoChromaFormat eChromaFormat)
 unsigned long GetNumDecodeSurfaces(cudaVideoCodec eCodec, unsigned int nWidth,
                                    unsigned int nHeight)
 {
+  const char *num_decode_surfaces_str = getenv("NV_NUM_DECODE_SURFACES");
+  if (num_decode_surfaces_str) {
+    return atoi(num_decode_surfaces_str);
+  }
+
   if (eCodec == cudaVideoCodec_VP9) {
     return 12;
   }
@@ -614,6 +619,10 @@ NvDecoder::NvDecoder(CUstream cuStream, CUcontext cuContext,
                      cudaVideoCodec eCodec, bool bLowLatency, int maxWidth,
                      int maxHeight)
 {
+  int lowLatency = getenv("NV_LOW_LATENCY") ? atoi(getenv("NV_LOW_LATENCY")) : 0;
+  if (lowLatency&1) setEndOfPicture = true;
+  if (lowLatency&2) bLowLatency = true;
+
   p_impl = new NvDecoderImpl();
   p_impl->m_cuvidStream = cuStream;
   p_impl->m_cuContext = cuContext;
@@ -716,6 +725,7 @@ bool NvDecoder::DecodeLockSurface(Buffer const* encFrame,
       encFrame ? encFrame->GetDataAs<const unsigned char>() : nullptr;
   packet.payload_size = encFrame ? encFrame->GetRawMemSize() : 0U;
   packet.flags = CUVID_PKT_TIMESTAMP;
+  if (setEndOfPicture) packet.flags |= CUVID_PKT_ENDOFPICTURE;
   packet.timestamp = pdata.pts;
   if (!decCtx.no_eos &&
       (nullptr == packet.payload || 0 == packet.payload_size)) {