diff --git a/miniz_oxide/src/deflate/core.rs b/miniz_oxide/src/deflate/core.rs
index 0064a51..9e9830d 100644
--- a/miniz_oxide/src/deflate/core.rs
+++ b/miniz_oxide/src/deflate/core.rs
@@ -1681,6 +1681,16 @@ pub(crate) fn flush_block(
         output.bit_buffer = d.params.saved_bit_buffer;
         output.bits_in = d.params.saved_bits_in;
 
+        // If we are at the start of the stream, write the zlib header
+        // if requested.  Note: Even if block-writing is skipped
+        // below, `block_index` is still incremented, so this is done
+        // only once
+        if d.params.flags & TDEFL_WRITE_ZLIB_HEADER != 0 && d.params.block_index == 0 {
+            let header = zlib::header_from_flags(d.params.flags);
+            output.put_bits_no_flush(header[0].into(), 8);
+            output.put_bits(header[1].into(), 8);
+        }
+
         if d.lz.total_bytes > 0 || flush == TDEFLFlush::Finish {
             // TODO: Don't think this second condition should be here but need to verify.
             let use_raw_block = (d.params.flags & TDEFL_FORCE_ALL_RAW_BLOCKS != 0)
@@ -1696,13 +1706,6 @@ pub(crate) fn flush_block(
 
             d.lz.init_flag();
 
-            // If we are at the start of the stream, write the zlib header if requested.
-            if d.params.flags & TDEFL_WRITE_ZLIB_HEADER != 0 && d.params.block_index == 0 {
-                let header = zlib::header_from_flags(d.params.flags);
-                output.put_bits_no_flush(header[0].into(), 8);
-                output.put_bits(header[1].into(), 8);
-            }
-
             // Output the block header.
             output.put_bits((flush == TDEFLFlush::Finish) as u32, 1);
 
diff --git a/miniz_oxide/tests/flush.rs b/miniz_oxide/tests/flush.rs
index 795fb9e..af87be1 100644
--- a/miniz_oxide/tests/flush.rs
+++ b/miniz_oxide/tests/flush.rs
@@ -93,6 +93,21 @@ fn test_flush() {
     }
 }
 
+/// Test that a sync at start of a stream inserts the Zlib header
+/// correctly
+#[test]
+fn check_zlib_with_sync_at_start() {
+    // Test sync gets zlib header
+    compress(b"", &[TDEFLFlush::Sync]);
+    compress(b"", &[TDEFLFlush::Partial]);
+
+    // Test that second sync DOESN'T get zlib header
+    // (i.e. `block_index` handling is correct)
+    compress(b"", &[TDEFLFlush::Sync, TDEFLFlush::Sync]);
+    compress(b"", &[TDEFLFlush::Partial, TDEFLFlush::Partial]);
+    compress(b"", &[TDEFLFlush::NoSync, TDEFLFlush::Sync]);
+}
+
 // Low-quality RNG, copied from test.rs
 struct Rng(u64);