diff --git a/format/FormatMultiImage.py b/format/FormatMultiImage.py
index 0b5c69bf9..cea489dfb 100644
--- a/format/FormatMultiImage.py
+++ b/format/FormatMultiImage.py
@@ -201,6 +201,7 @@ def get_imageset(
         assert not (as_sequence and lazy), "No lazy support for sequences"
 
         if single_file_indices is not None:
+            assert len(single_file_indices)
             single_file_indices = flex.size_t(single_file_indices)
 
         # Create an imageset or sequence
@@ -228,29 +229,30 @@ def get_imageset(
                 indices=single_file_indices,
             )
 
+            if single_file_indices is None:
+                single_file_indices = range(format_instance.get_num_images())
+
             # If any are None then read from format
-            if [beam, detector, goniometer, scan].count(None) != 0:
+            if not all((beam, detector, goniometer, scan)):
 
                 # Get list of models
-                beam = []
-                detector = []
-                goniometer = []
-                scan = []
-                for i in range(format_instance.get_num_images()):
-                    beam.append(format_instance.get_beam(i))
-                    detector.append(format_instance.get_detector(i))
-                    goniometer.append(format_instance.get_goniometer(i))
-                    scan.append(format_instance.get_scan(i))
-
-            if single_file_indices is None:
-                single_file_indices = list(range(format_instance.get_num_images()))
+                num_images = format_instance.get_num_images()
+                beam = [None] * num_images
+                detector = [None] * num_images
+                goniometer = [None] * num_images
+                scan = [None] * num_images
+                for i in single_file_indices:
+                    beam[i] = format_instance.get_beam(i)
+                    detector[i] = format_instance.get_detector(i)
+                    goniometer[i] = format_instance.get_goniometer(i)
+                    scan[i] = format_instance.get_scan(i)
 
             # Set the list of models
-            for i in range(len(single_file_indices)):
-                iset.set_beam(beam[single_file_indices[i]], i)
-                iset.set_detector(detector[single_file_indices[i]], i)
-                iset.set_goniometer(goniometer[single_file_indices[i]], i)
-                iset.set_scan(scan[single_file_indices[i]], i)
+            for i, index in enumerate(single_file_indices):
+                iset.set_beam(beam[index], i)
+                iset.set_detector(detector[index], i)
+                iset.set_goniometer(goniometer[index], i)
+                iset.set_scan(scan[index], i)
 
         else:
 
diff --git a/newsfragments/210.feature b/newsfragments/210.feature
new file mode 100644
index 000000000..bc464847e
--- /dev/null
+++ b/newsfragments/210.feature
@@ -0,0 +1,3 @@
+FormatMultiImage: When constructing an imageset with the indices of some (not
+all) single images in the container, we skip reading models for the images that
+were not requested. In some cases this speeds up imageset construction by 8x.
diff --git a/tests/test_imageset.py b/tests/test_imageset.py
index 714392f60..2b39c4d0e 100644
--- a/tests/test_imageset.py
+++ b/tests/test_imageset.py
@@ -3,20 +3,44 @@
 import os
 from builtins import range
 
+import mock
 import pytest
 import six.moves.cPickle as pickle
 
 from scitbx.array_family import flex
 
+import dxtbx.format.FormatHDF5SaclaMPCCD
 import dxtbx.format.image
 import dxtbx.format.Registry
 import dxtbx.tests.imagelist
 from dxtbx.format.FormatCBFMiniPilatus import FormatCBFMiniPilatus as FormatClass
 from dxtbx.imageset import ExternalLookup, ImageSequence, ImageSetData, ImageSetFactory
 from dxtbx.model import Beam, Detector, Panel
+from dxtbx.model.beam import BeamFactory
 from dxtbx.model.experiment_list import ExperimentListFactory
 
 
+@pytest.mark.parametrize("indices,expected_call_count", ((None, 4), ([1], 2)))
+def test_single_file_indices(indices, expected_call_count, dials_regression):
+    def dummy_beam():
+        return BeamFactory.simple(1.0)
+
+    with mock.patch.object(
+        dxtbx.format.FormatHDF5SaclaMPCCD.FormatHDF5SaclaMPCCD,
+        "_beam",
+        side_effect=dummy_beam,
+    ) as obj:
+        filename = os.path.join(
+            dials_regression,
+            "image_examples",
+            "SACLA_MPCCD_Cheetah",
+            "run266702-0-subset.h5",
+        )
+        format_class = dxtbx.format.Registry.get_format_class_for_file(filename)
+        format_class.get_imageset([filename], single_file_indices=indices)
+        assert obj.call_count == expected_call_count
+
+
 @pytest.mark.parametrize(
     "image",
     dxtbx.tests.imagelist.smv_images