diff --git a/configs/sample_person_shoe_two_stage.json b/configs/sample_person_shoe_two_stage.json
index 98bc00c..80a8cf3 100644
--- a/configs/sample_person_shoe_two_stage.json
+++ b/configs/sample_person_shoe_two_stage.json
@@ -37,6 +37,9 @@
           "type": "ai_yolo",
           "role": "filter",
           "enable": true,
+          "use_rga": true,
+          "rga_gate": "person_shoe_two_stage",
+          "rga_max_inflight": 4,
           "infer_fps": 2,
           "model_path": "./models/yolov8n-640.rknn",
           "model_version": "v8",
diff --git a/configs/sample_person_shoe_two_stage_balanced.json b/configs/sample_person_shoe_two_stage_balanced.json
index e4627ad..158f59d 100644
--- a/configs/sample_person_shoe_two_stage_balanced.json
+++ b/configs/sample_person_shoe_two_stage_balanced.json
@@ -37,6 +37,9 @@
           "type": "ai_yolo",
           "role": "filter",
           "enable": true,
+          "use_rga": true,
+          "rga_gate": "person_shoe_two_stage_balanced",
+          "rga_max_inflight": 4,
           "infer_fps": 2,
           "model_path": "./models/yolov8n-640.rknn",
           "model_version": "v8",
diff --git a/configs/sample_person_shoe_two_stage_recall.json b/configs/sample_person_shoe_two_stage_recall.json
index bf58c30..020520c 100644
--- a/configs/sample_person_shoe_two_stage_recall.json
+++ b/configs/sample_person_shoe_two_stage_recall.json
@@ -37,6 +37,9 @@
           "type": "ai_yolo",
           "role": "filter",
           "enable": true,
+          "use_rga": true,
+          "rga_gate": "person_shoe_two_stage_recall",
+          "rga_max_inflight": 4,
           "infer_fps": 3,
           "model_path": "./models/yolov8n-640.rknn",
           "model_version": "v8",
diff --git a/plugins/ai_yolo/ai_yolo_node.cpp b/plugins/ai_yolo/ai_yolo_node.cpp
index f1fb476..ab57e26 100644
--- a/plugins/ai_yolo/ai_yolo_node.cpp
+++ b/plugins/ai_yolo/ai_yolo_node.cpp
@@ -10,6 +10,7 @@
 #include <thread>
 #include <vector>
 
+#include "hw/i_image_processor.h"
 #include "hw/i_infer_backend.h"
 #include "node.h"
 #include "utils/dma_alloc.h"
@@ -732,6 +733,7 @@ public:
             LogError("[ai_yolo] no infer backend for node " + id_);
             return false;
         }
+        image_processor_ = ctx.image_processor;
 
 #if defined(RK3588_ENABLE_RKNN)
         if (model_path_.empty()) {
@@ -895,7 +897,6 @@ private:
                 return;
             }
 
-            if (frame->DmaFd() >= 0) frame->SyncStart();
             const size_t input_row = static_cast<size_t>(model_input_w_) * 3;
             const size_t input_size = input_row * static_cast<size_t>(model_input_h_);
             resized_input_.resize(input_size);
@@ -906,12 +907,49 @@ private:
                            src + static_cast<size_t>(y) * static_cast<size_t>(src_stride),
                            input_row);
                 }
+            } else if (image_processor_ && w > 0 && h > 0) {
+                Frame src_frame;
+                src_frame.width = w;
+                src_frame.height = h;
+                src_frame.format = PixelFormat::RGB;
+                src_frame.stride = src_stride;
+                src_frame.data = const_cast<uint8_t*>(src);
+                src_frame.data_size = frame->data_size;
+                src_frame.plane_count = 1;
+                src_frame.planes[0] = {src_frame.data, src_stride, static_cast<int>(frame->data_size), 0};
+
+                Frame resized_frame;
+                resized_frame.width = model_input_w_;
+                resized_frame.height = model_input_h_;
+                resized_frame.format = PixelFormat::RGB;
+
+                Status st = image_processor_->Resize(src_frame, resized_frame);
+                if (st.Ok()) {
+                    const uint8_t* resized_data = resized_frame.planes[0].data ? resized_frame.planes[0].data : resized_frame.data;
+                    const int resized_stride = resized_frame.planes[0].stride > 0
+                                                   ? resized_frame.planes[0].stride
+                                                   : (resized_frame.stride > 0 ? resized_frame.stride : static_cast<int>(input_row));
+                    if (resized_data && resized_stride > 0) {
+                        for (int y = 0; y < model_input_h_; ++y) {
+                            memcpy(resized_input_.data() + static_cast<size_t>(y) * input_row,
+                                   resized_data + static_cast<size_t>(y) * static_cast<size_t>(resized_stride),
+                                   input_row);
+                        }
+                    } else {
+                        ResizeRgbBilinear(src, w, h, src_stride,
+                                          resized_input_.data(), model_input_w_, model_input_h_,
+                                          static_cast<int>(input_row));
+                    }
+                } else {
+                    ResizeRgbBilinear(src, w, h, src_stride,
+                                      resized_input_.data(), model_input_w_, model_input_h_,
+                                      static_cast<int>(input_row));
+                }
             } else {
                 ResizeRgbBilinear(src, w, h, src_stride,
                                   resized_input_.data(), model_input_w_, model_input_h_,
                                   static_cast<int>(input_row));
             }
-            if (frame->DmaFd() >= 0) frame->SyncEnd();
             input.data = resized_input_.data();
             input.size = input_size;
             input.width = model_input_w_;
@@ -1252,6 +1290,7 @@ private:
 
     std::shared_ptr<SpscQueue<FramePtr>> input_queue_;
     std::vector<std::shared_ptr<SpscQueue<FramePtr>>> output_queues_;
+    std::shared_ptr<IImageProcessor> image_processor_;
     std::shared_ptr<IInferBackend> infer_backend_;
     uint64_t processed_ = 0;
 
diff --git a/src/graph_manager.cpp b/src/graph_manager.cpp
index 6b81ef6..0de6f06 100644
--- a/src/graph_manager.cpp
+++ b/src/graph_manager.cpp
@@ -495,7 +495,7 @@ bool Graph::Build(const SimpleJson& graph_cfg, PluginLoader& loader, size_t defa
         entry.enabled = node_val.ValueOr<bool>("enable", true);
         entry.metrics = std::make_shared<NodeMetrics>();
         entry.context.infer_backend = infer_backend_;
-        if (entry.type == "preprocess" || entry.type == "ai_shoe_det") {
+        if (entry.type == "preprocess" || entry.type == "ai_shoe_det" || entry.type == "ai_yolo") {
             entry.context.image_processor = HwFactory::CreateImageProcessor(entry.config);
         }