diff --git a/zenovis/xinxinoptix/ChiefDesignerEXR.h b/zenovis/xinxinoptix/ChiefDesignerEXR.h
index d7c1835176..e72966418f 100644
--- a/zenovis/xinxinoptix/ChiefDesignerEXR.h
+++ b/zenovis/xinxinoptix/ChiefDesignerEXR.h
@@ -161,4 +161,49 @@ inline void SaveMultiLayerEXR(
     file.writePixels (height);
 }
 
+inline void SaveMultiLayerEXR_half(
+    std::vector<half*> pixels
+    , int width
+    , int height
+    , std::vector<std::string> channels
+    , const char* exrFilePath
+) {
+    using namespace Imath;
+    using namespace Imf;
+
+    Header header(width, height);
+    ChannelList channelList;
+
+    const char *std_suffix = "RGB";
+    for (auto channel: channels) {
+        for (int i = 0; i < 3; i++) {
+            std::string name = zeno::format("{}{}", channel, std_suffix[i]);
+            channelList.insert(name, Channel(HALF));
+        }
+    }
+
+    header.channels() = channelList;
+
+    OutputFile file (exrFilePath, header);
+    FrameBuffer frameBuffer;
+
+    std::vector<std::vector<half>> data;
+    for (half *rgb: pixels) {
+        std::vector<half> half_rgb(width * height * 3);
+        for (auto i = 0; i < half_rgb.size(); i++) {
+            half_rgb[i] = rgb[i];
+        }
+        data.push_back(std::move(half_rgb));
+    }
+
+    for (auto i = 0; i < channels.size(); i++) {
+        frameBuffer.insert (zeno::format("{}R", channels[i]), Slice ( HALF, (char*) &data[i][0], sizeof (half) * 3, sizeof (half) * width * 3));
+        frameBuffer.insert (zeno::format("{}G", channels[i]), Slice ( HALF, (char*) &data[i][1], sizeof (half) * 3, sizeof (half) * width * 3));
+        frameBuffer.insert (zeno::format("{}B", channels[i]), Slice ( HALF, (char*) &data[i][2], sizeof (half) * 3, sizeof (half) * width * 3));
+    }
+
+    file.setFrameBuffer (frameBuffer);
+    file.writePixels (height);
+}
+
 }
diff --git a/zenovis/xinxinoptix/optixPathTracer.cpp b/zenovis/xinxinoptix/optixPathTracer.cpp
index ee314e3efb..ad0d2ffcc8 100644
--- a/zenovis/xinxinoptix/optixPathTracer.cpp
+++ b/zenovis/xinxinoptix/optixPathTracer.cpp
@@ -3746,6 +3746,55 @@ std::vector<float> optixgetimg_extra2(std::string name, int w, int h) {
     }
     return tex_data;
 }
+
+std::vector<half> optixgetimg_extra3(std::string name, int w, int h) {
+    std::vector<half> tex_data(w * h * 3);
+    if (name == "diffuse") {
+        std::vector<float> temp_buffer(w * h * 3);
+        cudaMemcpy(temp_buffer.data(), (void*)state.accum_buffer_d.handle, sizeof(temp_buffer[0]) * temp_buffer.size(), cudaMemcpyDeviceToHost);
+        for (auto i = 0; i < temp_buffer.size(); i++) {
+            tex_data[i] = temp_buffer[i];
+        }
+    }
+    else if (name == "specular") {
+        std::vector<float> temp_buffer(w * h * 3);
+        cudaMemcpy(temp_buffer.data(), (void*)state.accum_buffer_s.handle, sizeof(temp_buffer[0]) * temp_buffer.size(), cudaMemcpyDeviceToHost);
+        for (auto i = 0; i < temp_buffer.size(); i++) {
+            tex_data[i] = temp_buffer[i];
+        }
+    }
+    else if (name == "transmit") {
+        std::vector<float> temp_buffer(w * h * 3);
+        cudaMemcpy(temp_buffer.data(), (void*)state.accum_buffer_t.handle, sizeof(temp_buffer[0]) * temp_buffer.size(), cudaMemcpyDeviceToHost);
+        for (auto i = 0; i < temp_buffer.size(); i++) {
+            tex_data[i] = temp_buffer[i];
+        }
+    }
+    else if (name == "background") {
+        std::vector<half> temp_buffer(w * h);
+        cudaMemcpy(temp_buffer.data(), (void*)state.accum_buffer_b.handle, sizeof(temp_buffer[0]) * temp_buffer.size(), cudaMemcpyDeviceToHost);
+        for (auto i = 0; i < temp_buffer.size(); i++) {
+            tex_data[i * 3 + 0] = temp_buffer[i];
+            tex_data[i * 3 + 1] = temp_buffer[i];
+            tex_data[i * 3 + 2] = temp_buffer[i];
+        }
+    }
+    else if (name == "mask") {
+        cudaMemcpy(tex_data.data(), (void*)state.accum_buffer_m.handle, sizeof(half) * tex_data.size(), cudaMemcpyDeviceToHost);
+    }
+    else if (name == "color") {
+        std::vector<float> temp_buffer(w * h * 3);
+        cudaMemcpy(temp_buffer.data(), (void*)state.accum_buffer_p.handle, sizeof(temp_buffer[0]) * temp_buffer.size(), cudaMemcpyDeviceToHost);
+        for (auto i = 0; i < temp_buffer.size(); i++) {
+            tex_data[i] = temp_buffer[i];
+        }
+    }
+    else {
+        throw std::runtime_error("invalid optixgetimg_extra name: " + name);
+    }
+    zeno::image_flip_vertical((ushort3*)tex_data.data(), w, h);
+    return tex_data;
+}
 static void save_exr(float3* ptr, int w, int h, std::string path) {
     std::vector<float3> data(w * h);
     std::copy_n(ptr, w * h, data.data());
@@ -3826,14 +3875,14 @@ void optixrender(int fbo, int samples, bool denoise, bool simpleRender) {
         if (enable_output_aov) {
             if (enable_output_exr) {
                 zeno::create_directories_when_write_file(exr_path);
-                SaveMultiLayerEXR(
+                SaveMultiLayerEXR_half(
                         {
-                                optixgetimg_extra2("color", w, h).data(),
-                                optixgetimg_extra2("diffuse", w, h).data(),
-                                optixgetimg_extra2("specular", w, h).data(),
-                                optixgetimg_extra2("transmit", w, h).data(),
-                                optixgetimg_extra2("background", w, h).data(),
-                                optixgetimg_extra2("mask", w, h).data(),
+                                optixgetimg_extra3("color", w, h).data(),
+                                optixgetimg_extra3("diffuse", w, h).data(),
+                                optixgetimg_extra3("specular", w, h).data(),
+                                optixgetimg_extra3("transmit", w, h).data(),
+                                optixgetimg_extra3("background", w, h).data(),
+                                optixgetimg_extra3("mask", w, h).data(),
                         },
                         w,
                         h,