diff --git a/app/assets/filters/core/descratch.json b/app/assets/filters/core/descratch.json
index c993fc3..228c86a 100644
--- a/app/assets/filters/core/descratch.json
+++ b/app/assets/filters/core/descratch.json
@@ -7,6 +7,7 @@
   "category": "cleanup",
   "icon": "healing",
   "order": 2,
+  "maxBitDepth": 8,
 
   "dependencies": {
     "vs_plugins": ["libdescratch.dll"]
diff --git a/app/lib/models/filter_schema.dart b/app/lib/models/filter_schema.dart
index 29bb6c7..6f13c5d 100644
--- a/app/lib/models/filter_schema.dart
+++ b/app/lib/models/filter_schema.dart
@@ -357,6 +357,12 @@ class FilterSchema {
   /// Sort order in filter list.
   final int order;
 
+  /// Highest per-component bit depth this filter processes natively. When set
+  /// and the source exceeds it, the worker down-converts to this depth for the
+  /// pass (a lossy round-trip), so the UI warns of reduced precision. Null (the
+  /// default) means the filter handles the source's bit depth without loss.
+  final int? maxBitDepth;
+
   /// Dependencies required by this filter.
   final FilterDependencies? dependencies;
 
@@ -391,6 +397,7 @@ class FilterSchema {
     this.category,
     this.icon,
     this.order = 0,
+    this.maxBitDepth,
     this.dependencies,
     required this.methods,
     required this.parameters,
diff --git a/app/lib/utils/pixel_format.dart b/app/lib/utils/pixel_format.dart
new file mode 100644
index 0000000..0106c46
--- /dev/null
+++ b/app/lib/utils/pixel_format.dart
@@ -0,0 +1,67 @@
+// Utilities for interpreting FFmpeg pixel-format strings (e.g. the `pix_fmt`
+// reported by ffprobe: "yuv420p", "yuv422p10le", "yuv420p16le").
+
+/// Best-effort per-component bit depth for an FFmpeg pixel-format string.
+///
+/// Returns 8 for the common 8-bit formats (yuv420p, yuv422p, nv12, rgb24, …),
+/// and the explicit depth for higher-precision formats (yuv422p10le → 10,
+/// yuv420p16le → 16, p010le → 10, rgb48le → 16, gray12le → 12).
+///
+/// Unknown or null formats fall back to 8 — the conservative choice, so an
+/// unrecognized format never produces a spurious "quality will be reduced"
+/// warning.
+int pixelFormatBitDepth(String? pixFmt) {
+  if (pixFmt == null || pixFmt.trim().isEmpty) return 8;
+  final f = pixFmt.trim().toLowerCase();
+
+  // Planar YUV/GBR/gray with an explicit per-component depth suffix:
+  // yuv422p10le, yuv420p16le, gbrp12le, gray10le, gray16be, …
+  final planar = RegExp(r'(?:p|gray|gbrp?)(\d{1,2})(?:le|be)?$').firstMatch(f);
+  if (planar != null) return int.parse(planar.group(1)!);
+
+  // Semi-planar high-depth: p010le, p016le, p210le, p410le, p216le, …
+  // (the middle digit is the subsampling, the last two are the depth).
+  final semi = RegExp(r'^p\d(10|12|16)(?:le|be)?$').firstMatch(f);
+  if (semi != null) return int.parse(semi.group(1)!);
+
+  // Packed high-depth RGB: rgb48/bgr48 and rgba64/bgra64 are 16-bit/component.
+  if (RegExp(r'(?:48|64)(?:le|be)?$').hasMatch(f) &&
+      RegExp(r'^[argb]{3,4}').hasMatch(f)) {
+    return 16;
+  }
+
+  // Everything else (yuv420p, yuv422p, yuv411p, nv12, gray, rgb24, …) is 8-bit.
+  return 8;
+}
+
+/// Warning message when an enabled filter with a native [maxBitDepth] ceiling
+/// will down-convert a higher-bit-depth [pixelFormat] source (a lossy 8-bit
+/// round-trip), or null when no warning is warranted (pass off, no ceiling,
+/// unknown format, or the source already fits).
+String? filterBitDepthWarning({
+  required String filterName,
+  required bool enabled,
+  int? maxBitDepth,
+  String? pixelFormat,
+}) {
+  if (!enabled || maxBitDepth == null || pixelFormat == null) return null;
+  final depth = pixelFormatBitDepth(pixelFormat);
+  if (depth <= maxBitDepth) return null;
+  return '$filterName only processes in $maxBitDepth-bit. Your $depth-bit '
+      'source will be converted to $maxBitDepth-bit for this pass and back, '
+      'reducing colour precision.';
+}
+
+/// Warning message when converting chroma subsampling will also drop a
+/// higher-bit-depth [pixelFormat] source to 8-bit on output, or null when not
+/// applicable (not [converting], unknown format, or the source is 8-bit).
+String? chromaConversionBitDepthWarning({
+  required bool converting,
+  String? pixelFormat,
+}) {
+  if (!converting || pixelFormat == null) return null;
+  final depth = pixelFormatBitDepth(pixelFormat);
+  if (depth <= 8) return null;
+  return 'Your $depth-bit source will be output as 8-bit when converting '
+      'chroma subsampling. Choose "Original" to keep the source\'s bit depth.';
+}
diff --git a/app/lib/views/pass_settings/pass_settings_container.dart b/app/lib/views/pass_settings/pass_settings_container.dart
index 53791d5..4ff9d9c 100644
--- a/app/lib/views/pass_settings/pass_settings_container.dart
+++ b/app/lib/views/pass_settings/pass_settings_container.dart
@@ -3,9 +3,12 @@ import 'package:provider/provider.dart';
 
 import '../../models/dynamic_parameters.dart';
 import '../../models/filter_registry.dart';
+import '../../models/filter_schema.dart';
 import '../../models/processing_pipeline.dart';
 import '../../services/whisper_addon_manager.dart';
+import '../../utils/pixel_format.dart';
 import '../../viewmodels/main_viewmodel.dart';
+import '../../widgets/warning_banner.dart';
 import '../settings/dynamic_filter_panel.dart';
 import '../whisper_download_dialog.dart';
 
@@ -69,6 +72,7 @@ class PassSettingsContainer extends StatelessWidget {
               crossAxisAlignment: CrossAxisAlignment.stretch,
               children: [
                 _buildOpenCLWarning(context, viewModel, filterId, params),
+                _buildBitDepthWarning(context, viewModel, schema, params),
                 DynamicFilterPanelCompact(
                   schema: schema,
                   params: params,
@@ -113,35 +117,28 @@ class PassSettingsContainer extends StatelessWidget {
         : 'No OpenCL device detected. OpenCL acceleration will fall back to '
             'CPU NNEDI3.';
 
-    return Padding(
-      padding: const EdgeInsets.only(bottom: 12),
-      child: Container(
-        padding: const EdgeInsets.all(12),
-        decoration: BoxDecoration(
-          color: Theme.of(context).colorScheme.errorContainer,
-          borderRadius: BorderRadius.circular(8),
-        ),
-        child: Row(
-          crossAxisAlignment: CrossAxisAlignment.start,
-          children: [
-            Icon(
-              Icons.warning_amber_rounded,
-              size: 20,
-              color: Theme.of(context).colorScheme.onErrorContainer,
-            ),
-            const SizedBox(width: 8),
-            Expanded(
-              child: Text(
-                message,
-                style: Theme.of(context).textTheme.bodySmall?.copyWith(
-                      color: Theme.of(context).colorScheme.onErrorContainer,
-                    ),
-              ),
-            ),
-          ],
-        ),
-      ),
+    return WarningBanner(message: message);
+  }
+
+  /// Warning shown when an enabled filter can't process the source's bit depth
+  /// natively (schema `maxBitDepth`), so the worker down-converts to that depth
+  /// for the pass and restores it afterward — a lossy round-trip. E.g. DeScratch
+  /// is 8-bit only, so a 10-bit source loses precision through that pass.
+  /// Returns an empty widget when the pass is off or the source fits natively.
+  Widget _buildBitDepthWarning(
+    BuildContext context,
+    MainViewModel viewModel,
+    FilterSchema schema,
+    DynamicParameters params,
+  ) {
+    final message = filterBitDepthWarning(
+      filterName: schema.name,
+      enabled: params.enabled,
+      maxBitDepth: schema.maxBitDepth,
+      pixelFormat: viewModel.videoInfo?.pixelFormat,
     );
+    if (message == null) return const SizedBox.shrink();
+    return WarningBanner(message: message);
   }
 
   /// Handle parameter changes, with special model-download check for subtitles.
diff --git a/app/lib/views/settings/settings_dialog.dart b/app/lib/views/settings/settings_dialog.dart
index af09408..cc050c8 100644
--- a/app/lib/views/settings/settings_dialog.dart
+++ b/app/lib/views/settings/settings_dialog.dart
@@ -10,7 +10,9 @@ import '../../models/video_job.dart';
 import '../../services/dependency_manager.dart';
 import '../../services/hardware_encoder_detector.dart';
 import '../../services/update_checker.dart';
+import '../../utils/pixel_format.dart';
 import '../../viewmodels/main_viewmodel.dart';
+import '../../widgets/warning_banner.dart';
 
 class SettingsDialog extends StatefulWidget {
   const SettingsDialog({super.key});
@@ -713,6 +715,7 @@ class _OutputSettingsTabState extends State<_OutputSettingsTab> {
                               .withValues(alpha: 0.6),
                         ),
                   ),
+                  ..._buildChromaBitDepthWarning(viewModel, settings),
                 ],
               ),
             ),
@@ -758,6 +761,23 @@ class _OutputSettingsTabState extends State<_OutputSettingsTab> {
     );
   }
 
+  /// Warning shown when converting chroma subsampling would also drop a
+  /// higher-bit-depth source to 8-bit on output. A chroma-subsampling
+  /// conversion forces an 8-bit output format (YUV420P8/YUV422P8), whereas
+  /// "Original" preserves the source format and its bit depth. Returns an empty
+  /// list when nothing needs saying (source is 8-bit, unknown, or "Original").
+  List<Widget> _buildChromaBitDepthWarning(
+    MainViewModel viewModel,
+    EncodingSettings settings,
+  ) {
+    final message = chromaConversionBitDepthWarning(
+      converting: settings.chromaSubsampling != ChromaSubsampling.original,
+      pixelFormat: viewModel.videoInfo?.pixelFormat,
+    );
+    if (message == null) return const [];
+    return [const SizedBox(height: 12), WarningBanner(message: message)];
+  }
+
   /// Whether a hardware encoder is relevant to the current platform's GPU APIs:
   /// VideoToolbox is macOS-only; QSV/NVENC/AMF apply to Windows and Linux.
   /// (Software/ProRes/lossless codecs are platform-agnostic.)
diff --git a/app/lib/widgets/warning_banner.dart b/app/lib/widgets/warning_banner.dart
new file mode 100644
index 0000000..2f42a56
--- /dev/null
+++ b/app/lib/widgets/warning_banner.dart
@@ -0,0 +1,42 @@
+import 'package:flutter/material.dart';
+
+/// A compact, non-blocking advisory banner: a warning icon plus a message in
+/// the theme's error-container colours. Used for "this will still work, but…"
+/// notices such as missing OpenCL devices or bit-depth quality reductions.
+class WarningBanner extends StatelessWidget {
+  const WarningBanner({super.key, required this.message});
+
+  final String message;
+
+  @override
+  Widget build(BuildContext context) {
+    final scheme = Theme.of(context).colorScheme;
+    return Padding(
+      padding: const EdgeInsets.only(bottom: 12),
+      child: Container(
+        padding: const EdgeInsets.all(12),
+        decoration: BoxDecoration(
+          color: scheme.errorContainer,
+          borderRadius: BorderRadius.circular(8),
+        ),
+        child: Row(
+          crossAxisAlignment: CrossAxisAlignment.start,
+          children: [
+            Icon(Icons.warning_amber_rounded,
+                size: 20, color: scheme.onErrorContainer),
+            const SizedBox(width: 8),
+            Expanded(
+              child: Text(
+                message,
+                style: Theme.of(context)
+                    .textTheme
+                    .bodySmall
+                    ?.copyWith(color: scheme.onErrorContainer),
+              ),
+            ),
+          ],
+        ),
+      ),
+    );
+  }
+}
diff --git a/app/test/pixel_format_test.dart b/app/test/pixel_format_test.dart
new file mode 100644
index 0000000..3a9cbfd
--- /dev/null
+++ b/app/test/pixel_format_test.dart
@@ -0,0 +1,130 @@
+import 'package:flutter_test/flutter_test.dart';
+import 'package:vapourbox/utils/pixel_format.dart';
+
+void main() {
+  group('pixelFormatBitDepth', () {
+    test('8-bit formats', () {
+      for (final f in [
+        'yuv420p', 'yuv422p', 'yuv444p', 'yuv411p', 'yuv440p',
+        'yuvj420p', 'yuvj422p', 'nv12', 'nv21', 'gray', 'rgb24', 'bgr24',
+        'rgba', 'bgra', 'gbrp',
+      ]) {
+        expect(pixelFormatBitDepth(f), 8, reason: f);
+      }
+    });
+
+    test('10-bit formats', () {
+      for (final f in [
+        'yuv420p10le', 'yuv422p10le', 'yuv444p10le', 'yuv420p10be',
+        'gbrp10le', 'gray10le', 'p010le', 'p210le', 'p410le',
+      ]) {
+        expect(pixelFormatBitDepth(f), 10, reason: f);
+      }
+    });
+
+    test('12-bit formats', () {
+      for (final f in ['yuv420p12le', 'yuv444p12le', 'gray12le', 'gbrp12le']) {
+        expect(pixelFormatBitDepth(f), 12, reason: f);
+      }
+    });
+
+    test('16-bit formats', () {
+      for (final f in [
+        'yuv420p16le', 'yuv444p16le', 'gray16le', 'gbrp16le',
+        'p016le', 'p216le', 'rgb48le', 'bgr48le', 'rgba64le', 'bgra64be',
+      ]) {
+        expect(pixelFormatBitDepth(f), 16, reason: f);
+      }
+    });
+
+    test('null / empty / unknown fall back to 8 (no spurious warning)', () {
+      expect(pixelFormatBitDepth(null), 8);
+      expect(pixelFormatBitDepth(''), 8);
+      expect(pixelFormatBitDepth('  '), 8);
+      expect(pixelFormatBitDepth('unknown'), 8);
+      expect(pixelFormatBitDepth('some_future_format'), 8);
+    });
+
+    test('case and whitespace tolerant', () {
+      expect(pixelFormatBitDepth('YUV422P10LE'), 10);
+      expect(pixelFormatBitDepth(' yuv420p '), 8);
+    });
+
+    test('the ProRes 422 case from the bug report', () {
+      // The 10-bit source that triggered the IVTC fix.
+      expect(pixelFormatBitDepth('yuv422p10le'), 10);
+    });
+  });
+
+  group('filterBitDepthWarning', () {
+    String? call({
+      bool enabled = true,
+      int? maxBitDepth = 8,
+      String? pixelFormat = 'yuv422p10le',
+    }) =>
+        filterBitDepthWarning(
+          filterName: 'DeScratch',
+          enabled: enabled,
+          maxBitDepth: maxBitDepth,
+          pixelFormat: pixelFormat,
+        );
+
+    test('warns for a 10-bit source on an 8-bit-only filter', () {
+      final msg = call();
+      expect(msg, isNotNull);
+      expect(msg, contains('DeScratch'));
+      expect(msg, contains('8-bit'));
+      expect(msg, contains('10-bit'));
+    });
+
+    test('silent when the pass is disabled', () {
+      expect(call(enabled: false), isNull);
+    });
+
+    test('silent when the filter has no bit-depth ceiling', () {
+      expect(call(maxBitDepth: null), isNull);
+    });
+
+    test('silent for an 8-bit source', () {
+      expect(call(pixelFormat: 'yuv420p'), isNull);
+    });
+
+    test('silent when the source fits (depth == ceiling)', () {
+      expect(call(maxBitDepth: 10, pixelFormat: 'yuv422p10le'), isNull);
+    });
+
+    test('silent when the pixel format is unknown/null', () {
+      expect(call(pixelFormat: null), isNull);
+    });
+  });
+
+  group('chromaConversionBitDepthWarning', () {
+    test('warns when converting a >8-bit source', () {
+      final msg = chromaConversionBitDepthWarning(
+          converting: true, pixelFormat: 'yuv422p10le');
+      expect(msg, isNotNull);
+      expect(msg, contains('10-bit'));
+      expect(msg, contains('8-bit'));
+    });
+
+    test('silent when keeping Original (not converting)', () {
+      expect(
+          chromaConversionBitDepthWarning(
+              converting: false, pixelFormat: 'yuv422p10le'),
+          isNull);
+    });
+
+    test('silent for an 8-bit source even when converting', () {
+      expect(
+          chromaConversionBitDepthWarning(
+              converting: true, pixelFormat: 'yuv420p'),
+          isNull);
+    });
+
+    test('silent when the pixel format is null', () {
+      expect(
+          chromaConversionBitDepthWarning(converting: true, pixelFormat: null),
+          isNull);
+    });
+  });
+}