diff --git a/src/d3d11/d3d11_initializer.cpp b/src/d3d11/d3d11_initializer.cpp
index 7a264e38..ac0a7b03 100644
--- a/src/d3d11/d3d11_initializer.cpp
+++ b/src/d3d11/d3d11_initializer.cpp
@@ -9,7 +9,8 @@ namespace dxvk {
           D3D11Device*                pParent)
   : m_parent(pParent),
     m_device(pParent->GetDXVKDevice()),
-    m_context(m_device->createContext(DxvkContextType::Supplementary)) {
+    m_context(m_device->createContext(DxvkContextType::Supplementary)),
+    m_stagingBuffer(m_device, StagingBufferSize) {
     m_context->beginRecording(
       m_device->createCommandList());
   }
@@ -81,20 +82,22 @@ namespace dxvk {
     const D3D11_SUBRESOURCE_DATA*     pInitialData) {
     std::lock_guard<dxvk::mutex> lock(m_mutex);
 
-    DxvkBufferSlice bufferSlice = pBuffer->GetBufferSlice();
+    Rc<DxvkBuffer> buffer = pBuffer->GetBuffer();
 
     if (pInitialData != nullptr && pInitialData->pSysMem != nullptr) {
-      m_transferMemory   += bufferSlice.length();
+      auto stagingSlice = m_stagingBuffer.alloc(buffer->info().size);
+      std::memcpy(stagingSlice.mapPtr(0), pInitialData->pSysMem, stagingSlice.length());
+
+      m_transferMemory += buffer->info().size;
       m_transferCommands += 1;
       
-      m_context->uploadBuffer(
-        bufferSlice.buffer(),
-        pInitialData->pSysMem);
+      m_context->uploadBuffer(buffer,
+        stagingSlice.buffer(),
+        stagingSlice.offset());
     } else {
       m_transferCommands += 1;
 
-      m_context->initBuffer(
-        bufferSlice.buffer());
+      m_context->initBuffer(buffer);
     }
 
     FlushImplicit();
@@ -284,6 +287,8 @@ namespace dxvk {
     
     m_transferCommands = 0;
     m_transferMemory   = 0;
+
+    m_stagingBuffer.reset();
   }
 
 
diff --git a/src/d3d11/d3d11_initializer.h b/src/d3d11/d3d11_initializer.h
index fd3d4dfa..732b2058 100644
--- a/src/d3d11/d3d11_initializer.h
+++ b/src/d3d11/d3d11_initializer.h
@@ -1,5 +1,7 @@
 #pragma once
 
+#include "../dxvk/dxvk_staging.h"
+
 #include "d3d11_buffer.h"
 #include "d3d11_texture.h"
 
@@ -18,6 +20,9 @@ namespace dxvk {
   class D3D11Initializer {
     constexpr static size_t MaxTransferMemory    = 32 * 1024 * 1024;
     constexpr static size_t MaxTransferCommands  = 512;
+
+    // Use a staging buffer with a linear allocator to service small uploads
+    constexpr static VkDeviceSize StagingBufferSize = 1ull << 20;
   public:
 
     D3D11Initializer(
@@ -45,6 +50,8 @@ namespace dxvk {
     D3D11Device*      m_parent;
     Rc<DxvkDevice>    m_device;
     Rc<DxvkContext>   m_context;
+    
+    DxvkStagingBuffer m_stagingBuffer;
 
     size_t            m_transferCommands  = 0;
     size_t            m_transferMemory    = 0;
diff --git a/src/dxvk/dxvk_context.cpp b/src/dxvk/dxvk_context.cpp
index 3bd1d98a..706979fc 100644
--- a/src/dxvk/dxvk_context.cpp
+++ b/src/dxvk/dxvk_context.cpp
@@ -2537,20 +2537,18 @@ namespace dxvk {
 
   void DxvkContext::uploadBuffer(
     const Rc<DxvkBuffer>&           buffer,
-    const void*                     data) {
+    const Rc<DxvkBuffer>&           source,
+          VkDeviceSize              sourceOffset) {
     auto bufferSlice = buffer->getSliceHandle();
-
-    auto stagingSlice = m_staging.alloc(bufferSlice.length);
-    auto stagingHandle = stagingSlice.getSliceHandle();
-    std::memcpy(stagingHandle.mapPtr, data, bufferSlice.length);
+    auto sourceSlice = source->getSliceHandle(sourceOffset, buffer->info().size);
 
     VkBufferCopy2 copyRegion = { VK_STRUCTURE_TYPE_BUFFER_COPY_2 };
-    copyRegion.srcOffset = stagingHandle.offset;
+    copyRegion.srcOffset = sourceSlice.offset;
     copyRegion.dstOffset = bufferSlice.offset;
     copyRegion.size      = bufferSlice.length;
 
     VkCopyBufferInfo2 copyInfo = { VK_STRUCTURE_TYPE_COPY_BUFFER_INFO_2 };
-    copyInfo.srcBuffer = stagingHandle.handle;
+    copyInfo.srcBuffer = sourceSlice.handle;
     copyInfo.dstBuffer = bufferSlice.handle;
     copyInfo.regionCount = 1;
     copyInfo.pRegions = &copyRegion;
@@ -2574,7 +2572,7 @@ namespace dxvk {
         buffer->info().stages, buffer->info().access);
     }
 
-    m_cmd->trackResource<DxvkAccess::Read>(stagingSlice.buffer());
+    m_cmd->trackResource<DxvkAccess::Read>(source);
     m_cmd->trackResource<DxvkAccess::Write>(buffer);
   }
 
diff --git a/src/dxvk/dxvk_context.h b/src/dxvk/dxvk_context.h
index ce936ce1..a4c5fa89 100644
--- a/src/dxvk/dxvk_context.h
+++ b/src/dxvk/dxvk_context.h
@@ -1151,14 +1151,17 @@ namespace dxvk {
     
     /**
      * \brief Uses transfer queue to initialize buffer
-     * 
-     * Only safe to use if the buffer is not in use by the GPU.
+     *
+     * Always replaces the entire buffer. Only safe to use
+     * if the buffer is currently not in use by the GPU.
      * \param [in] buffer The buffer to initialize
-     * \param [in] data The data to copy to the buffer
+     * \param [in] source Staging buffer containing data
+     * \param [in] sourceOffset Offset into staging buffer
      */
     void uploadBuffer(
       const Rc<DxvkBuffer>&           buffer,
-      const void*                     data);
+      const Rc<DxvkBuffer>&           source,
+            VkDeviceSize              sourceOffset);
     
     /**
      * \brief Uses transfer queue to initialize image