diff --git a/tests/ImagePreprocessorGPUFusedTest.cpp b/tests/ImagePreprocessorGPUFusedTest.cpp
index 0400591d..fcd122b8 100644
--- a/tests/ImagePreprocessorGPUFusedTest.cpp
+++ b/tests/ImagePreprocessorGPUFusedTest.cpp
@@ -69,10 +69,13 @@ void CheckFusedMatchesCPU(CompressedImageMode mode, T err_value, uint32_t seed)
 
     PixelMask mask(x);
     // Mask a deterministic scatter of pixels, so masked-vs-error-vs-saturated priority is exercised
-    // rather than assumed.
-    auto &m = const_cast<std::vector<uint32_t> &>(mask.GetMask());
-    for (size_t i = 0; i < npixels; i += 997) m[i] = 1;
-    for (size_t i = 13; i < npixels; i += 4001) m[i] = 1;
+    // rather than assumed. It goes in through LoadUserMask because the mask derives the
+    // byte-per-pixel form the GPU preprocessor uploads: writing the bitfield behind its back leaves
+    // that form stale, and the device then works from a mask the CPU reference does not have.
+    std::vector<uint32_t> user_mask(mask.GetMask().size(), 0);
+    for (size_t i = 0; i < npixels; i += 997) user_mask[i] = 1;
+    for (size_t i = 13; i < npixels; i += 4001) user_mask[i] = 1;
+    mask.LoadUserMask(x, user_mask);
 
     const auto img = MakeImage<T>(npixels, err_value, seed);
     JFJochBitShuffleCompressor compressor(CompressionAlgorithm::BSHUF_LZ4);
@@ -132,8 +135,9 @@ TEST_CASE("ImagePreprocessorGPU_FusedMatchesHostUpload", "[ImagePreprocessorGPU]
     DiffractionExperiment x = MakeExperiment(32000);
     const size_t npixels = x.GetPixelsNum();
     PixelMask mask(x);
-    auto &m = const_cast<std::vector<uint32_t> &>(mask.GetMask());
-    for (size_t i = 0; i < npixels; i += 1301) m[i] = 1;
+    std::vector<uint32_t> user_mask(mask.GetMask().size(), 0);
+    for (size_t i = 0; i < npixels; i += 1301) user_mask[i] = 1;
+    mask.LoadUserMask(x, user_mask);
 
     const auto img = MakeImage<uint32_t>(npixels, UINT32_MAX, 77);
     JFJochBitShuffleCompressor compressor(CompressionAlgorithm::BSHUF_LZ4);