diff --git a/requirements-linux-lite.txt b/requirements-linux-lite.txt
index 8192b73a..28d00cc6 100644
--- a/requirements-linux-lite.txt
+++ b/requirements-linux-lite.txt
@@ -20,6 +20,6 @@ torch==2.5.1
 torchvision==0.20.1
 pymediainfo
 torchmetrics
-celux==0.5.3
+celux==0.5.4
 https://github.com/NevermindNilas/TAS-Modes-Host/releases/download/main/upscale_ncnn_py-1.2.0-cp312-none-manylinux1_x86_64.whl
 https://github.com/NevermindNilas/TAS-Modes-Host/releases/download/main/rife_ncnn_vulkan_python-1.3.0-cp312-cp312-linux_x86_64.whl
diff --git a/requirements-linux.txt b/requirements-linux.txt
index 95395c51..bec1e10e 100644
--- a/requirements-linux.txt
+++ b/requirements-linux.txt
@@ -25,6 +25,6 @@ torch==2.5.1
 torchvision==0.20.1
 pymediainfo
 torchmetrics
-celux==0.5.3
+celux==0.5.4
 https://github.com/NevermindNilas/TAS-Modes-Host/releases/download/main/upscale_ncnn_py-1.2.0-cp312-none-manylinux1_x86_64.whl
 https://github.com/NevermindNilas/TAS-Modes-Host/releases/download/main/rife_ncnn_vulkan_python-1.3.0-cp312-cp312-linux_x86_64.whl
diff --git a/requirements-windows-lite.txt b/requirements-windows-lite.txt
index 69e187bc..feba77be 100644
--- a/requirements-windows-lite.txt
+++ b/requirements-windows-lite.txt
@@ -17,7 +17,7 @@ rich
 rich-argparse
 flask
 torchmetrics
-celux==0.5.3
+celux==0.5.4
 torch==2.5.1
 pymediainfo
 torchvision==0.20.1
diff --git a/requirements-windows.txt b/requirements-windows.txt
index 34f3f183..df5b9ce7 100644
--- a/requirements-windows.txt
+++ b/requirements-windows.txt
@@ -26,6 +26,6 @@ tensorrt-cu12-bindings==10.7.0
 tensorrt-cu12_libs==10.7.0
 torchmetrics
 pymediainfo
-celux==0.5.3
+celux==0.5.4
 https://github.com/NevermindNilas/TAS-Modes-Host/releases/download/main/rife_ncnn_vulkan_python-1.2.1-cp312-cp312-win_amd64.whl
 https://github.com/NevermindNilas/TAS-Modes-Host/releases/download/main/upscale_ncnn_py-1.2.0-cp312-none-win_amd64.whl
diff --git a/src/dedup/dedup.py b/src/dedup/dedup.py
index 9e544af3..db6b13a4 100644
--- a/src/dedup/dedup.py
+++ b/src/dedup/dedup.py
@@ -299,7 +299,9 @@ def __call__(self, frame):
 
         with torch.cuda.stream(self.normStream):
             frame = self.prepareFrame(frame)
-            self.dummyInput.copy_(torch.cat((self.prevFrame, frame), dim=1))
+            self.dummyInput.copy_(
+                torch.cat((self.prevFrame, frame), dim=1), non_blocking=True
+            )
         self.normStream.synchronize()
 
         with torch.cuda.stream(self.stream):
diff --git a/src/utils/ffmpegSettings.py b/src/utils/ffmpegSettings.py
index eca3e65a..85b94bef 100644
--- a/src/utils/ffmpegSettings.py
+++ b/src/utils/ffmpegSettings.py
@@ -385,7 +385,6 @@ def matchEncoder(encode_method: str):
     return command
 
 
-
 class BuildBuffer:
     def __init__(
         self,
@@ -400,7 +399,7 @@ def __init__(
         resizeMethod: str = "bicubic",
         width: int = 1920,
         height: int = 1080,
-        mainPath: str = ""
+        mainPath: str = "",
     ):
         """
         Initializes the BuildBuffer class.
@@ -445,22 +444,31 @@ def __init__(
             try:
                 if outpoint != 0.0:
                     self.reader = VideoReader(
-                        videoInput, device="cpu", num_threads=decodeThreads, filters=filters
+                        videoInput,
+                        device="cpu",
+                        num_threads=decodeThreads,
+                        filters=filters,
                     )([float(inpoint), float(outpoint)])
                 else:
                     self.reader = VideoReader(
-                        videoInput, device="cpu", num_threads=decodeThreads, filters=filters
+                        videoInput,
+                        device="cpu",
+                        num_threads=decodeThreads,
+                        filters=filters,
                     )
+                logging.info("Using celux.VideoReader for video decoding")
             except Exception as e:
                 logging.error(f"Failed to initialize celux.VideoReader: {e}")
                 logging.info("Falling back to OpenCV for video decoding")
                 self.useOpenCV = True
                 self.initializeOpenCV(videoInput, inputFramePoint, outputFramePoint)
-        
+
         # Delete from memory, can't trust the garbage collector
         del jsonMetadata
 
-    def initializeOpenCV(self, videoInput: str, inputFramePoint: int = 0, outputFramePoint: int = 0):
+    def initializeOpenCV(
+        self, videoInput: str, inputFramePoint: int = 0, outputFramePoint: int = 0
+    ):
         """
         Initializes the OpenCV video reader.
         """
@@ -567,6 +575,8 @@ def isQueueEmpty(self) -> bool:
             Whether the decoding buffer is empty.
         """
         return self.decodeBuffer.empty()
+
+
 class WriteBuffer:
     def __init__(
         self,