Cycles: load 8 bit image textures as half float for some color spaces

For non-raw, non-sRGB color spaces, always use half float even if that uses
more memory. Otherwise the precision loss from conversion to scene linear or
sRGB (as natively understood by the texture sampling) can be too much.

This also required a change to do alpha association ourselves instead of OIIO,
because in OIIO alpha multiplication happens before conversion to half float
and that gives too much precision loss.

Ref T68926

1 files changed, 2 insertions, 7 deletions

diff --git a/intern/cycles/scene/image.cpp b/intern/cycles/scene/image.cpp
index 2aa9a6bc1a1..1b44162351a 100644
--- a/intern/cycles/scene/image.cpp
+++ b/intern/cycles/scene/image.cpp
@@ -272,17 +272,12 @@ void ImageMetaData::detect_colorspace()
     compress_as_srgb = true;
   }
   else {
-    /* Always compress non-raw 8bit images as scene linear + sRGB, as a
-     * heuristic to keep memory usage the same without too much data loss
-     * due to quantization in common cases. */
-    compress_as_srgb = (type == IMAGE_DATA_TYPE_BYTE || type == IMAGE_DATA_TYPE_BYTE4);
-
     /* If colorspace conversion needed, use half instead of short so we can
      * represent HDR values that might result from conversion. */
-    if (type == IMAGE_DATA_TYPE_USHORT) {
+    if (type == IMAGE_DATA_TYPE_BYTE || type == IMAGE_DATA_TYPE_USHORT) {
       type = IMAGE_DATA_TYPE_HALF;
     }
-    else if (type == IMAGE_DATA_TYPE_USHORT4) {
+    else if (type == IMAGE_DATA_TYPE_BYTE4 || type == IMAGE_DATA_TYPE_USHORT4) {
       type = IMAGE_DATA_TYPE_HALF4;
     }
   }