diff --git a/image_analysis/image_preprocessing/ImagePreprocessorGPU.cu b/image_analysis/image_preprocessing/ImagePreprocessorGPU.cu
index 643058cb6..a40ff94bb 100644
--- a/image_analysis/image_preprocessing/ImagePreprocessorGPU.cu
+++ b/image_analysis/image_preprocessing/ImagePreprocessorGPU.cu
@@ -352,7 +352,16 @@ void ImagePreprocessorGPU::PinInputRegion(const void *ptr, size_t bytes) {
         cuda_err(cudaHostUnregister(const_cast<void *>(pinned_input)));
     pinned_input = nullptr;
     pinned_input_bytes = 0;
-    cuda_err(cudaHostRegister(const_cast<void *>(ptr), bytes, cudaHostRegisterDefault));
+    // The buffer may already be page-locked by whoever owns it - the receiver registers its image
+    // buffer once and hands out ranges inside it - and registering a range twice is an error rather
+    // than a no-op. There is nothing to do in that case: the upload is already a DMA, and the
+    // region is not ours to unregister later, so it is deliberately not remembered.
+    const auto err = cudaHostRegister(const_cast<void *>(ptr), bytes, cudaHostRegisterDefault);
+    if (err == cudaErrorHostMemoryAlreadyRegistered) {
+        cudaGetLastError();
+        return;
+    }
+    cuda_err(err);
     pinned_input = ptr;
     pinned_input_bytes = bytes;
 }