diff --git a/IPython/nbformat/v2/nbjson.py b/IPython/nbformat/v2/nbjson.py
index 249439f..b86297d 100644
--- a/IPython/nbformat/v2/nbjson.py
+++ b/IPython/nbformat/v2/nbjson.py
@@ -16,9 +16,8 @@ Authors:
 # Imports
 #-----------------------------------------------------------------------------
 
-from base64 import encodestring
 from .nbbase import from_dict
-from .rwbase import NotebookReader, NotebookWriter, base64_decode
+from .rwbase import NotebookReader, NotebookWriter, restore_bytes
 import json
 
 #-----------------------------------------------------------------------------
@@ -26,9 +25,10 @@ import json
 #-----------------------------------------------------------------------------
 
 class BytesEncoder(json.JSONEncoder):
+    """A JSON encoder that accepts b64 (and other *ascii*) bytestrings."""
     def default(self, obj):
         if isinstance(obj, bytes):
-            return encodestring(obj).decode('ascii')
+            return obj.decode('ascii')
         return json.JSONEncoder.default(self, obj)
 
 
@@ -40,7 +40,7 @@ class JSONReader(NotebookReader):
         return nb
 
     def to_notebook(self, d, **kwargs):
-        return base64_decode(from_dict(d))
+        return restore_bytes(from_dict(d))
 
 
 class JSONWriter(NotebookWriter):
diff --git a/IPython/nbformat/v2/rwbase.py b/IPython/nbformat/v2/rwbase.py
index 2641639..cbcb50a 100644
--- a/IPython/nbformat/v2/rwbase.py
+++ b/IPython/nbformat/v2/rwbase.py
@@ -19,31 +19,67 @@ Authors:
 from base64 import encodestring, decodestring
 import pprint
 
+from IPython.utils.py3compat import str_to_bytes
+
 #-----------------------------------------------------------------------------
 # Code
 #-----------------------------------------------------------------------------
 
+def restore_bytes(nb):
+    """Restore bytes of image data from unicode-only formats.
+    
+    Base64 encoding is handled elsewhere.  Bytes objects in the notebook are
+    always b64-encoded. We DO NOT encode/decode around file formats.
+    """
+    for ws in nb.worksheets:
+        for cell in ws.cells:
+            if cell.cell_type == 'code':
+                for output in cell.outputs:
+                    if 'png' in output:
+                        output.png = str_to_bytes(output.png, 'ascii')
+                    if 'jpeg' in output:
+                        output.jpeg = str_to_bytes(output.jpeg, 'ascii')
+    return nb
+
+
+# b64 encode/decode are never actually used, because all bytes objects in
+# the notebook are already b64-encoded, and we don't need/want to double-encode
+
 def base64_decode(nb):
-    """Base64 encode all bytes objects in the notebook."""
+    """Restore all bytes objects in the notebook from base64-encoded strings.
+    
+    Note: This is never used
+    """
     for ws in nb.worksheets:
         for cell in ws.cells:
             if cell.cell_type == 'code':
-                if 'png' in cell:
-                    cell.png = bytes(decodestring(cell.png))
-                if 'jpeg' in cell:
-                    cell.jpeg = bytes(decodestring(cell.jpeg))
+                for output in cell.outputs:
+                    if 'png' in output:
+                        if isinstance(output.png, unicode):
+                            output.png = output.png.encode('ascii')
+                        output.png = decodestring(output.png)
+                    if 'jpeg' in output:
+                        if isinstance(output.jpeg, unicode):
+                            output.jpeg = output.jpeg.encode('ascii')
+                        output.jpeg = decodestring(output.jpeg)
     return nb
 
 
 def base64_encode(nb):
-    """Base64 decode all binary objects in the notebook."""
+    """Base64 encode all bytes objects in the notebook.
+    
+    These will be b64-encoded unicode strings
+    
+    Note: This is never used
+    """
     for ws in nb.worksheets:
         for cell in ws.cells:
             if cell.cell_type == 'code':
-                if 'png' in cell:
-                    cell.png = unicode(encodestring(cell.png))
-                if 'jpeg' in cell:
-                    cell.jpeg = unicode(encodestring(cell.jpeg))
+                for output in cell.outputs:
+                    if 'png' in output:
+                        output.png = encodestring(output.png).decode('ascii')
+                    if 'jpeg' in output:
+                        output.jpeg = encodestring(output.jpeg).decode('ascii')
     return nb
 
 
diff --git a/IPython/nbformat/v2/tests/nbexamples.py b/IPython/nbformat/v2/tests/nbexamples.py
index b567884..2f6174b 100644
--- a/IPython/nbformat/v2/tests/nbexamples.py
+++ b/IPython/nbformat/v2/tests/nbexamples.py
@@ -1,10 +1,15 @@
+import os
+from base64 import encodestring
+
 from ..nbbase import (
     NotebookNode,
     new_code_cell, new_text_cell, new_worksheet, new_notebook, new_output,
     new_metadata, new_author
 )
 
-
+# some random base64-encoded *bytes*
+png = encodestring(os.urandom(5))
+jpeg = encodestring(os.urandom(6))
 
 ws = new_worksheet(name='worksheet1')
 
@@ -42,8 +47,8 @@ ws.cells.append(new_code_cell(
         output_text=u'<array a>',
         output_html=u'The HTML rep',
         output_latex=u'$a$',
-        output_png=b'data',
-        output_jpeg=b'data',
+        output_png=png,
+        output_jpeg=jpeg,
         output_svg=u'<svg>',
         output_json=u'json data',
         output_javascript=u'var i=0;',
@@ -53,8 +58,8 @@ ws.cells.append(new_code_cell(
         output_text=u'<array a>',
         output_html=u'The HTML rep',
         output_latex=u'$a$',
-        output_png=b'data',
-        output_jpeg=b'data',
+        output_png=png,
+        output_jpeg=jpeg,
         output_svg=u'<svg>',
         output_json=u'json data',
         output_javascript=u'var i=0;'