diff --git a/src/decoder/image.rs b/src/decoder/image.rs
index 2ffa77c..2b8c035 100644
--- a/src/decoder/image.rs
+++ b/src/decoder/image.rs
@@ -1204,7 +1204,7 @@ impl Image {
             let mut packed = vec![0u8; chunk_row_bytes];
 
             for row in buf.chunks_mut(layout.row_stride).take(data_dims.1 as usize) {
-                reader.read_samples(&mut packed)?;
+                reader.read_exact(&mut packed)?;
 
                 let out_row = &mut row[..data_row_bytes];
                 unpack_bits(&packed, out_row, tiff_bps, samples_per_data_row, out_bps);
@@ -1238,7 +1238,7 @@ impl Image {
         } else if is_output_chunk_rows && is_all_bits {
             // Here we can read directly into the output buffer itself.
             let tile = &mut buf[..chunk_row_bytes * data_dims.1 as usize];
-            reader.read_samples(tile)?;
+            reader.read_exact(tile)?;
 
             for row in tile.chunks_mut(chunk_row_bytes) {
                 super::fix_endianness_and_predict(
@@ -1272,7 +1272,7 @@ impl Image {
             // this case is handled specially when needed.
             let mut encoded = vec![0u8; chunk_row_bytes];
             for row in buf.chunks_mut(layout.row_stride).take(data_dims.1 as usize) {
-                reader.read_samples(&mut encoded)?;
+                reader.read_exact(&mut encoded)?;
 
                 let row = &mut row[..data_row_bytes];
                 match color_type.bit_depth() {
@@ -1313,7 +1313,7 @@ impl Image {
                     let skip = u64::try_from(chunk_row_bytes - data_row_bytes)?;
                     reader.read_padded_samples(&mut row[..used], skip)?;
                 } else {
-                    reader.read_samples(&mut row[..used])?;
+                    reader.read_exact(&mut row[..used])?;
                 }
 
                 super::fix_endianness_and_predict(
@@ -1371,7 +1371,7 @@ impl Image {
                 // width, and writing their horizontal padding would spill into the following
                 // output row (and predict/invert would run past the chunk's own columns).
                 let row = &mut row[..data_row_bytes];
-                reader.read_samples(&mut encoded)?;
+                reader.read_exact(&mut encoded)?;
 
                 Self::compact_photometric_bytes(&mut encoded, row, &photo_range);
 
diff --git a/src/decoder/stream.rs b/src/decoder/stream.rs
index 71e8937..326ba67 100644
--- a/src/decoder/stream.rs
+++ b/src/decoder/stream.rs
@@ -124,6 +124,9 @@ impl<R: Read> EndianReader<R> {
 
 /// A variant of `io::Read` that takes one block to the end.
 ///
+/// This is derived from `Read` so that any special optimization done by `std` (or any nightly crate
+/// for that matter) surely remains usable.
+///
 /// We want our decompression engines to be much more specialized than the standard library trait.
 /// Firstly there is more information available to communicate (the tiff Type, known decompressed
 /// size, …) and secondly they can potentially *reset*. This could be solved with an extension trait
@@ -134,9 +137,7 @@ impl<R: Read> EndianReader<R> {
 /// discarding data *after* it is put into the output buffer. Yet, for some decompression that write
 /// alone is considerable work where it could simply skip bits more efficiently. So the reader would
 /// likely want to know about a sample mask.
-pub trait ReadSamples {
-    fn read_samples(&mut self, buf: &mut [u8]) -> io::Result<()>;
-
+pub trait ReadSamples: Read {
     /// Read a row of samples and throw away some additional padding samples at the end.
     ///
     /// This is automatically implemented for `Self: Read` as a copy into `io::Sink`.
@@ -144,10 +145,6 @@ pub trait ReadSamples {
 }
 
 impl<R: Read> ReadSamples for R {
-    fn read_samples(&mut self, buf: &mut [u8]) -> io::Result<()> {
-        self.read_exact(buf)
-    }
-
     fn read_padded_samples(&mut self, buf: &mut [u8], discard: u64) -> io::Result<()> {
         self.read_exact(buf)?;
         std::io::copy(&mut self.by_ref().take(discard), &mut std::io::sink())?;