diff --git a/transcode_bench.py b/transcode_bench.py
index ff7b773..fb72bb2 100755
--- a/transcode_bench.py
+++ b/transcode_bench.py
@@ -358,21 +358,19 @@ class TranscodeJob:
             # Build video filter chain for realistic transcoding
             vf_filters = []
 
-            # For VA-API, we need to scale and upload to hardware
-            if 'vaapi' in self.encoder:
-                if self.hw_decode:
-                    # HW decode: download from GPU, scale on CPU, upload back
-                    # (GPU scaling not widely supported, so use CPU)
-                    vf_filters.append('hwdownload')
-                    vf_filters.append('format=nv12')
-
-                # Scale on CPU (works everywhere)
-                vf_filters.append('scale=1280x720')
+            # Handle decode format conversion and scaling
+            if self.hw_decode:
+                # HW decode: download from GPU to CPU first
+                vf_filters.append('hwdownload')
                 vf_filters.append('format=nv12')
+
+            # Scale on CPU (works everywhere)
+            vf_filters.append('scale=1280x720')
+            vf_filters.append('format=nv12')
+
+            # Upload to GPU only if HEVC encoder is hardware-accelerated
+            if 'vaapi' in self.hevc_encoder:
                 vf_filters.append('hwupload')
-            else:
-                # Other encoders: just scale
-                vf_filters.append('scale=1280x720')
 
             if vf_filters:
                 cmd.extend(['-vf', ','.join(vf_filters)])