Comment tweaks

Author: scotts

src/torchcodec/_core/CpuDeviceInterface.cpp

@@ -70,12 +70,6 @@ void CpuDeviceInterface::initializeVideo(
   // If we have any transforms, replace filters_ with the filter strings from
   // the transforms. As noted above, we decide between swscale and filtergraph
   // when we actually decode a frame.
-  //
-  // Note: We explicitly add the format conversion filter at the end to ensure
-  // that color conversion happens AFTER the transforms, not before. This
-  // matches the behavior of the reference generation in the test suite.
-  // Without this, FFmpeg's automatic format negotiation might insert the
-  // conversion before the transforms, which would produce different results.
   std::stringstream filters;
   bool first = true;
   for (const auto& transform : transforms) {

src/torchcodec/_core/CpuDeviceInterface.h

@@ -109,9 +109,8 @@ class CpuDeviceInterface : public DeviceInterface {
   //
   // If we had to apply transforms in the output pixel format and colorspace,
   // we could achieve that by calling sws_scale() twice: once to do the resize
-  // and another time to do the format conversion. But that goes against the
-  // whole point of calling sws_scale() directly, as it's a performance
-  // optimization.
+  // and another time to do the format conversion. But that will be slower,
+  // which goes against the whole point of calling sws_scale() directly.
   std::string filters_ = "format=rgb24";
 
   // The flags we supply to swsContext_, if it used. The flags control the

test/generate_reference_resources.py

@@ -51,10 +51,11 @@ def generate_frame_by_index(
     )
     output_bmp = f"{base_path}.bmp"
 
-    # Note that we have an exlicit format conversion to rgb24 in our filtergraph specification,
-    # which always happens AFTER any of the filters that we receive as input. We do this to
-    # ensure that the color conversion happens AFTER the filters, matching the behavior of the
-    # torchcodec filtergraph implementation.
+    # Note that we have an exlicit format conversion to rgb24 in our filtergraph
+    # specification, and we always place the user-supplied filters BEFORE the
+    # format conversion. We do this to ensure that the filters are applied in
+    # the pixel format and colorspace of the input frames. This behavior matches
+    # the TorchCodec implementation.
     select = f"select='eq(n\\,{frame_index})'"
     format = "format=rgb24"
     if filters is not None: