diff --git a/CHANGELOG.md b/CHANGELOG.md
index 9ceb4a6..1a6088d 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -13,6 +13,7 @@
 - Bookmark→library save and bookmark Save As now plumb the active site's `CategoryFetcher` through to the filename template, so `%artist%`/`%character%` tokens render correctly instead of silently dropping out when saving a post that wasn't previewed first
 - Info panel no longer silently drops tags that failed to land in a cached category — any tag from `post.tag_list` not rendered under a known category section now appears in an "Other" bucket, so partial cache coverage can't make individual tags invisible
 - `BooruClient._request` retries now cover `httpx.RemoteProtocolError` and `httpx.ReadError` in addition to the existing timeout/connect/network set — an overloaded booru that drops the TCP connection mid-response no longer fails the whole search on the first try
+- VRAM retained when no video is playing — `stop()` now frees the GL render context (textures + FBOs) instead of just dropping the hwdec surface pool. Context is recreated lazily on next `play_file()` via `ensure_gl_init()` (~5ms, invisible behind network fetch)
 
 ### Refactored
 - `category_fetcher` batch tag-API params are now built by a shared `_build_tag_api_params` helper instead of duplicated across `fetch_via_tag_api` and `_probe_batch_api`
diff --git a/booru_viewer/gui/media/mpv_gl.py b/booru_viewer/gui/media/mpv_gl.py
index 9927325..c50b2dc 100644
--- a/booru_viewer/gui/media/mpv_gl.py
+++ b/booru_viewer/gui/media/mpv_gl.py
@@ -111,7 +111,20 @@ class _MpvGLWidget(QWidget):
             self._gl.makeCurrent()
             self._init_gl()
 
-    def cleanup(self) -> None:
+    def release_render_context(self) -> None:
+        """Free the GL render context without terminating mpv.
+
+        Releases all GPU-side textures and FBOs that the render context
+        holds. The next ``ensure_gl_init()`` call (from ``play_file``)
+        recreates the context cheaply (~5ms). This is the difference
+        between "mpv is idle but holding VRAM" and "mpv is idle and
+        clean."
+
+        Safe to call when mpv has no active file (after
+        ``mpv.command('stop')``). After this, ``_paint_gl`` is a no-op
+        (``_ctx is None`` guard) and mpv won't fire frame-ready
+        callbacks because there's no render context to trigger them.
+        """
         if self._ctx:
             # GL context must be current so mpv can release its textures
             # and FBOs on the correct context. Without this, drivers that
@@ -123,6 +136,10 @@ class _MpvGLWidget(QWidget):
             finally:
                 self._gl.doneCurrent()
             self._ctx = None
+            self._gl_inited = False
+
+    def cleanup(self) -> None:
+        self.release_render_context()
         if self._mpv:
             self._mpv.terminate()
             self._mpv = None
diff --git a/booru_viewer/gui/media/video_player.py b/booru_viewer/gui/media/video_player.py
index d4c0c51..cc7efa4 100644
--- a/booru_viewer/gui/media/video_player.py
+++ b/booru_viewer/gui/media/video_player.py
@@ -491,6 +491,11 @@ class VideoPlayer(QWidget):
                 # teardown and rejects the write, GL context destruction
                 # still drops the surface pool eventually.
                 pass
+        # Free the GL render context so its internal textures and FBOs
+        # release VRAM while no video is playing. The next play_file()
+        # call recreates the context via ensure_gl_init() (~5ms cost,
+        # swamped by the network fetch for uncached videos).
+        self._gl_widget.release_render_context()
         self._time_label.setText("0:00")
         self._duration_label.setText("0:00")
         self._seek_slider.setRange(0, 0)