diff --git a/intern/cycles/device/cuda/device_impl.cpp b/intern/cycles/device/cuda/device_impl.cpp
index efe4b659132..7b388c465c9 100644
--- a/intern/cycles/device/cuda/device_impl.cpp
+++ b/intern/cycles/device/cuda/device_impl.cpp
@@ -733,6 +733,15 @@ void CUDADevice::tex_alloc(device_texture &mem)
   }
 
   /* Image Texture Storage */
+  /* Cycles expects to read all texture data as normalized float values in
+   * kernel/device/gpu/image.h. But storing all data as floats would be very inefficient due to the
+   * huge size of float textures. So in the code below, we define different texture types including
+   * integer types, with the aim of using CUDA's default promotion behavior of integer data to
+   * floating point data in the range [0, 1], as noted in the CUDA documentation on
+   * cuTexObjectCreate API Call.
+   * Note that 32-bit integers are not supported by this promotion behavior and cannot be used
+   * with Cycles's current implementation in kernel/device/gpu/image.h.
+   */
   CUarray_format_enum format;
   switch (mem.data_type) {
     case TYPE_UCHAR:
@@ -741,12 +750,6 @@ void CUDADevice::tex_alloc(device_texture &mem)
     case TYPE_UINT16:
       format = CU_AD_FORMAT_UNSIGNED_INT16;
       break;
-    case TYPE_UINT:
-      format = CU_AD_FORMAT_UNSIGNED_INT32;
-      break;
-    case TYPE_INT:
-      format = CU_AD_FORMAT_SIGNED_INT32;
-      break;
     case TYPE_FLOAT:
       format = CU_AD_FORMAT_FLOAT;
       break;
@@ -900,6 +903,8 @@ void CUDADevice::tex_alloc(device_texture &mem)
     texDesc.addressMode[1] = address_mode;
     texDesc.addressMode[2] = address_mode;
     texDesc.filterMode = filter_mode;
+    /* CUDA's flag CU_TRSF_READ_AS_INTEGER is intentionally not used and it is
+     * significant, see above an explanation about how Blender treat textures. */
     texDesc.flags = CU_TRSF_NORMALIZED_COORDINATES;
 
     thread_scoped_lock lock(device_mem_map_mutex);