diff --git a/docker-compose.nvidia.yml b/docker-compose.nvidia.yml
new file mode 100644
index 00000000..44a35494
--- /dev/null
+++ b/docker-compose.nvidia.yml
@@ -0,0 +1,12 @@
+version: "3.8"
+
+services:
+  transcoder:
+    runtime: nvidia
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - capabilities: [gpu]
+    environment:
+      - GOTRANSCODER_HWACCEL=nvidia
diff --git a/transcoder/Dockerfile b/transcoder/Dockerfile
index 95f62aeb..a4882850 100644
--- a/transcoder/Dockerfile
+++ b/transcoder/Dockerfile
@@ -2,6 +2,7 @@ FROM golang:1.21 as build
 RUN apt-get update \
 	&& apt-get install --no-install-recommends --no-install-suggests -y \
 	ffmpeg libavformat-dev libavutil-dev libswscale-dev libmediainfo-dev \
+	&& apt-get clean autoclean -y \
 	&& apt-get autoremove -y
 WORKDIR /app
 COPY go.mod go.sum ./
diff --git a/transcoder/Dockerfile.dev b/transcoder/Dockerfile.dev
index 843877ce..47acc5eb 100644
--- a/transcoder/Dockerfile.dev
+++ b/transcoder/Dockerfile.dev
@@ -2,6 +2,7 @@ FROM golang:1.21
 RUN apt-get update \
 	&& apt-get install --no-install-recommends --no-install-suggests -y \
 	ffmpeg libavformat-dev libavutil-dev libswscale-dev libmediainfo-dev \
+	&& apt-get clean autoclean -y \
 	&& apt-get autoremove -y
 RUN go install github.com/bokwoon95/wgo@latest
 WORKDIR /app
diff --git a/transcoder/src/hwaccel.go b/transcoder/src/hwaccel.go
index 0405be6d..20e5a9d5 100644
--- a/transcoder/src/hwaccel.go
+++ b/transcoder/src/hwaccel.go
@@ -1,7 +1,11 @@
 package src
 
+import "log"
+
 func DetectHardwareAccel() HwAccelT {
-	name := "disabled"
+	name := GetEnvOr("GOTRANSCODER_HWACCEL", "disabled")
+	log.Printf("Using hardware acceleration: %s", name)
+
 	switch name {
 	case "nvidia":
 		return HwAccelT{
@@ -17,6 +21,7 @@ func DetectHardwareAccel() HwAccelT {
 				"-c:v", "h264_nvenc",
 				"-preset", "fast",
 			},
+			ScaleFilter: "hwupload_cuda,scale_cuda=%d:%d:force_original_aspect_ratio=decrease",
 		}
 	default:
 		return HwAccelT{
@@ -27,6 +32,7 @@ func DetectHardwareAccel() HwAccelT {
 				// superfast or ultrafast would produce a file extremly big so we prever veryfast or faster.
 				"-preset", "faster",
 			},
+			ScaleFilter: "scale=%d:%d:force_original_aspect_ratio=decrease",
 		}
 	}
 }
diff --git a/transcoder/src/settings.go b/transcoder/src/settings.go
index c54573e3..c6af123d 100644
--- a/transcoder/src/settings.go
+++ b/transcoder/src/settings.go
@@ -20,6 +20,7 @@ type HwAccelT struct {
 	Name        string
 	DecodeFlags []string
 	EncodeFlags []string
+	ScaleFilter string
 }
 
 var Settings = SettingsT{
diff --git a/transcoder/src/videostream.go b/transcoder/src/videostream.go
index af476f22..5a936639 100644
--- a/transcoder/src/videostream.go
+++ b/transcoder/src/videostream.go
@@ -29,6 +29,16 @@ func (vs *VideoStream) getOutPath(encoder_id int) string {
 	return fmt.Sprintf("%s/segment-%s-%d-%%d.ts", vs.file.Out, vs.quality, encoder_id)
 }
 
+func closestMultiple(n int32, x int32) int32 {
+	if x > n {
+		return x
+	}
+
+	n = n + x/2
+	n = n - (n % x)
+	return n
+}
+
 func (vs *VideoStream) getTranscodeArgs(segments string) []string {
 	args := []string{
 		"-map", "0:V:0",
@@ -42,9 +52,11 @@ func (vs *VideoStream) getTranscodeArgs(segments string) []string {
 	}
 
 	args = append(args, Settings.HwAccel.EncodeFlags...)
+	width := int32(float64(vs.quality.Height()) / float64(vs.file.Info.Video.Height) * float64(vs.file.Info.Video.Width))
+	// force a width that is a multiple of two else some apps behave badly.
+	width = closestMultiple(width, 2)
 	args = append(args,
-		// resize but keep aspect ratio (also force a width that is a multiple of two else some apps behave badly.
-		"-vf", fmt.Sprintf("scale=-2:'min(%d,ih)'", vs.quality.Height()),
+		"-vf", fmt.Sprintf(Settings.HwAccel.ScaleFilter, width, vs.quality.Height()),
 		// Even less sure but bufsize are 5x the avergae bitrate since the average bitrate is only
 		// useful for hls segments.
 		"-bufsize", fmt.Sprint(vs.quality.MaxBitrate()*5),