diff --git a/src/plugins/intel_gna/src/debug_new_pass.cpp b/src/plugins/intel_gna/src/debug_new_pass.cpp
index 4bdd50b4e6a9fe..33e3679284a8fb 100644
--- a/src/plugins/intel_gna/src/debug_new_pass.cpp
+++ b/src/plugins/intel_gna/src/debug_new_pass.cpp
@@ -17,10 +17,13 @@ void DebugVisualize(ov::pass::Manager& manager, const std::string& name) {
 #ifdef DEBUG_VISUALIZE
     static unsigned counter = 0;
     std::stringstream ss;
+#ifdef DEBUG_VISUALIZETREE
     ss << counter << name << ".png";
     manager.register_pass<ov::pass::VisualizeTree>(ss.str());
-    //ss << counter << name;
-    //manager.register_pass<ov::pass::Serialize>(ss.str() + ".xml", ss.str() + ".bin");
+#else
+    ss << counter << name;
+    manager.register_pass<ov::pass::Serialize>(ss.str() + ".xml", ss.str() + ".bin");
+#endif
     ++counter;
 #endif
 }
diff --git a/src/plugins/intel_gna/src/debug_new_pass.hpp b/src/plugins/intel_gna/src/debug_new_pass.hpp
index 9bb85cf509c0b0..c75cd4ad566607 100644
--- a/src/plugins/intel_gna/src/debug_new_pass.hpp
+++ b/src/plugins/intel_gna/src/debug_new_pass.hpp
@@ -11,6 +11,8 @@
 
 #undef DEBUG_VISUALIZE
 //#define DEBUG_VISUALIZE 1
+#undef DEBUG_VISUALIZETREE
+//#define DEBUG_VISUALIZETREE 1
 
 #define EMUTEX_DEBUG_CHECKPOINT std::cout << "[EMUTEX DEBUG] CHECKPOINT " << __FILE__ << ":" << __LINE__ << std::endl;
 #define EMUTEX_DEBUG_CHECKPOINT_MESSAGE(message) std::cout << "[EMUTEX DEBUG] CHECKPOINT " << __FILE__ << ":" << __LINE__ << \
diff --git a/src/plugins/intel_gna/src/gna_transformations_pipeline.cpp b/src/plugins/intel_gna/src/gna_transformations_pipeline.cpp
index 9ec84366eca7c2..8f0b80b253bcf2 100644
--- a/src/plugins/intel_gna/src/gna_transformations_pipeline.cpp
+++ b/src/plugins/intel_gna/src/gna_transformations_pipeline.cpp
@@ -80,6 +80,7 @@ void TransformationsPipeline::apply(const std::shared_ptr<ov::Model>& model,
     // In OV API 2.0(IRv10) default convertion to fp32 (inputs, outputs and weights) is disabled
     // and we need to run the ConvertPrecision transformation to support old networks.
     manager.register_pass<ov::pass::ConvertPrecision>(precisions_array{{ngraph::element::f16, ngraph::element::f32}});
+    intel_gna_debug::DebugVisualize(manager, "start");
     manager.register_pass<ov::pass::ConvertMVN1ToMVN6>();
     manager.register_pass<ov::intel_gna::pass::DecomposeMVN>();
     manager.register_pass<ov::pass::CommonOptimizations>();
@@ -124,16 +125,17 @@ void TransformationsPipeline::apply(const std::shared_ptr<ov::Model>& model,
     manager.register_pass<ov::intel_gna::pass::InsertCopyBeforeLayerToBeEliminated>();
     // TODO enable this transformation for networks without convolutions
     if (ov::op::util::has_op_with_type<ngraph::opset7::Convolution>(model) || ov::op::util::has_op_with_type<ngraph::opset7::MaxPool>(model)) {
-    manager.register_pass<ov::intel_gna::pass::TransposeNCHW>();
-    manager.register_pass<ov::intel_gna::pass::ReshapeTransposeSubstitute>();
-    manager.register_pass<ov::pass::TransposeSinkingGeneral>();
-    manager.register_pass<ov::intel_gna::pass::GatherSinkingGeneral>();
-    manager.register_pass<ov::pass::ReshapeSequenceFusion>();
-    manager.register_pass<ov::pass::TransposeToReshape>();
-    manager.register_pass<ov::intel_gna::pass::GnaConvolutionFusion>();
+        manager.register_pass<ov::intel_gna::pass::TransposeNCHW>();
+        manager.register_pass<ov::intel_gna::pass::ReshapeTransposeSubstitute>();
+        manager.register_pass<ov::pass::TransposeSinkingGeneral>();
+        manager.register_pass<ov::intel_gna::pass::GatherSinkingGeneral>();
+        manager.register_pass<ov::pass::ReshapeSequenceFusion>();
+        manager.register_pass<ov::pass::TransposeToReshape>();
+        manager.register_pass<ov::intel_gna::pass::GnaConvolutionFusion>();
     }
     manager.register_pass<ov::intel_gna::pass::RemoveInputsProcessing>(subgraph_cpu_map);
     manager.register_pass<ov::intel_gna::pass::RemoveOutputsProcessing>(subgraph_cpu_map);
+    intel_gna_debug::DebugVisualize(manager, "after_our_transformations");
     manager.register_pass<ov::pass::ConvertOpSet3ToOpSet2>();
     manager.register_pass<ov::pass::ConvertOpSet2ToOpSet1>();
     manager.register_pass<ngraph::pass::ConvertOpSet1ToLegacy>();
@@ -195,6 +197,8 @@ void TransformationsPipeline::apply(const std::shared_ptr<ov::Model>& model,
     // Operations Max and Min aren't supported
     pass_config->disable<ov::pass::ConcatReduceFusion>();
 
+    intel_gna_debug::DebugVisualize(manager, "final");
+
     manager.run_passes(model);
 
     is_ngraph_passes_used = true;
diff --git a/src/plugins/intel_gna/src/transformations/gather_sinking_transpose_reshape.cpp b/src/plugins/intel_gna/src/transformations/gather_sinking_transpose_reshape.cpp
index 5f13a178cd68aa..bc867f2a75a434 100644
--- a/src/plugins/intel_gna/src/transformations/gather_sinking_transpose_reshape.cpp
+++ b/src/plugins/intel_gna/src/transformations/gather_sinking_transpose_reshape.cpp
@@ -142,6 +142,9 @@ bool AreFlattenShapes(const Shape& shape1, const Shape& shape2) {
     while (shape1[i] == shape2[i]) {
         ++i;
     }
+    // consider only last dimension to be flatten/unflatten
+    if (shape1.size() - 1 != i && shape2.size() - 1 != i)
+        return false;
     // min_shape.back() == MULTIPLY(max_shape.begin() + i, max_shape.end())
     const size_t mult1 = std::accumulate(shape1.begin() + i, shape1.end(), 1, std::multiplies<size_t>());
     const size_t mult2 = std::accumulate(shape2.begin() + i, shape2.end(), 1, std::multiplies<size_t>());