diff --git a/src/r36s_dlna_browser/player/gstreamer_backend.py b/src/r36s_dlna_browser/player/gstreamer_backend.py
index 3923fd7..110ca5d 100644
--- a/src/r36s_dlna_browser/player/gstreamer_backend.py
+++ b/src/r36s_dlna_browser/player/gstreamer_backend.py
@@ -431,26 +431,24 @@ class GStreamerBackend(PlayerBackend):
         # Hardware decode (NV12): insert a videoscale → capsfilter chain inside a
         # GstBin before the appsink so playbin accepts it as a single video-sink.
         #
-        # videoscale(method=nearest-neighbour, add-borders=True) scales the
-        # decoded source to fit within the video area while preserving the
-        # original aspect ratio.  Black NV12 borders fill any leftover space
-        # (letterbox / pillarbox), avoiding any stretch distortion.
+        # videoscale(method=nearest-neighbour) scales the decoded source to the
+        # video-area width while letting GStreamer pick the output height from
+        # the source's native aspect ratio.  Constraining only the width (not
+        # both width and height) means GStreamer will never stretch the video:
+        # it always preserves the source DAR.  The resulting frame is then
+        # centred in the SDL viewport by _fit_frame_to_viewport() which adds
+        # the necessary letterbox/pillarbox margins through SDL_RenderCopy.
         #   Nearest-neighbour skips ~56% of source rows so only ~44% of source
         #   cache lines are fetched; Python memmove drops from ~32 ms to ~1 ms.
         #
-        # capsfilter — enforces the output NV12 dimensions.
-        # Use the actual video area inside the HUD (full window minus margins)
-        # so the scale target matches the drawable region exactly.
-        # Dimensions are rounded down to even numbers (NV12 chroma subsampling
-        # requires both width and height to be divisible by 2).
+        # capsfilter — enforces format=NV12 and output width; height is left as
+        # a range so GStreamer can choose the correct value from the source AR.
         vp_w, vp_h, vp_top, vp_bottom, vp_left, vp_right = self._viewport
-        video_w = max(2, vp_w - vp_left - vp_right)
-        video_h = max(2, vp_h - vp_top - vp_bottom)
+        video_w = max(4, vp_w - vp_left - vp_right)
         scale_w = (video_w // 2) * 2
-        scale_h = (video_h // 2) * 2
-        if scale_w < 2 or scale_h < 2:
-            scale_w, scale_h = 640, 480
-        log.info("NV12 appsink: videoscale(nearest) → %dx%d before appsink", scale_w, scale_h)
+        if scale_w < 4:
+            scale_w = 640
+        log.info("NV12 appsink: videoscale(nearest) → width=%d (AR-preserving) before appsink", scale_w)
 
         scale      = self._gst.ElementFactory.make("videoscale",  "vscale")
         capsfilter = self._gst.ElementFactory.make("capsfilter",  "vcaps")
@@ -463,15 +461,13 @@ class GStreamerBackend(PlayerBackend):
 
         # nearest-neighbour: accesses only the source pixels needed for each
         # output sample (strided reads), skipping ~56% of source rows entirely.
-        # add-borders=True: letterbox/pillarbox to preserve the source aspect
-        # ratio instead of stretching to fill the target dimensions.
+        # Height is unconstrained (range 2–2160) so GStreamer picks whatever
+        # height preserves the source aspect ratio for the chosen width.
         scale.set_property("method", 0)
-        scale.set_property("add-borders", True)
         capsfilter.set_property(
             "caps",
             self._gst.Caps.from_string(
-                f"video/x-raw,format=NV12,width={scale_w},height={scale_h},"
-                f"pixel-aspect-ratio=1/1"
+                f"video/x-raw,format=NV12,width={scale_w},height=(int)[2,2160]"
             ),
         )