diff --git a/cuburn/genome/blend.py b/cuburn/genome/blend.py
index 6abd1d8..f9f4241 100644
--- a/cuburn/genome/blend.py
+++ b/cuburn/genome/blend.py
@@ -9,14 +9,73 @@ from itertools import izip_longest
 from scipy.ndimage.filters import gaussian_filter1d
 
 import spectypes
-import spec
-import util
-from util import get
+import specs
+from use import Wrapper
+from util import get, json_encode, resolve_spec
 import variations
 
-# TODO: move to better place before checkin
-default_blend_opts = {'nloops': 2, 'duration': 2, 'xform_sort': 'weightflip'}
+def node_to_anim(node, half):
+    if half:
+        osrc, odst = -0.25, 0.25
+    else:
+        osrc, odst = 0, 1
+    src = apply_temporal_offset(node, osrc)
+    dst = apply_temporal_offset(node, odst)
+    edge = dict(blend=dict(duration=odst-osrc, xform_sort='natural'))
+    return blend(src, dst, edge)
 
+def edge_to_anim(gdb, edge):
+    edge = resolve(gdb, edge)
+    src, osrc = _split_ref_id(edge.link.src)
+    dst, odst = _split_ref_id(edge.link.dst)
+    src = apply_temporal_offset(resolve(gdb, src), osrc)
+    dst = apply_temporal_offset(resolve(gdb, dst), odst)
+    return blend(src, dst, edit)
+
+def resolve(gdb, item):
+    """
+    Given an item, recursively retrieve its base items, then merge according
+    to type. Returns the merged dict.
+    """
+    is_edge = (item['type'] == 'edge')
+    spec = specs.toplevels[item['type']]
+    def go(i):
+        if i.get('base') is not None:
+            return go(gdb.get(i['base'])) + [i]
+        return [i]
+    items = map(flatten, go(item))
+    out = {}
+
+    for k in set(ik for i in items for ik in i):
+        sp = _resolve_spec(spec, k)
+        vs = [i.get(k) for i in items if k in i]
+        # TODO: dict and list negation; early-stage removal of negated knots?
+        if is_edge and isinstance(sp, (Spline, List)):
+            r = sum(vs, [])
+        else:
+            r = vs[-1]
+        out[k] = r
+    return unflatten(out)
+
+def _split_ref_id(s):
+    sp = s.split('@')
+    if len(sp) == 1:
+        return sp, 0
+    return sp[0], float(sp[1])
+
+def apply_temporal_offset(node, offset=0):
+    """
+    Given a ``node`` dict, return a node with all periodic splines rotated by
+    ``offset * velocity``, with the same velocity.
+    """
+    class TemporalOffsetWrapper(Wrapper):
+        def wrap_spline(self, path, spec, val):
+            if spec.period is not None and isinstance(val, list) and val[1]:
+                position, velocity = val
+                return [position + offset * velocity, velocity]
+            return val
+    wr = TemporalOffsetWrapper(node)
+    return wr.visit(wr)
 
 def blend(src, dst, edit={}):
     """
@@ -26,19 +85,22 @@ def blend(src, dst, edit={}):
     animation. These should be plain node dicts (hierarchical, pre-merged,
     and adjusted for loop temporal offset).
 
-    ``edit`` is an optional edit dict, also hierarchical and pre-merged.
+    ``edge`` is an edge dict, also hierarchical and pre-merged. (It can be
+    empty, in violation of the spec, to support rendering straight from nodes
+    without having to insert anything into the genome database.)
 
     Returns the animation spec as a plain dict.
     """
     # By design, the blend element will contain only scalar values (no
     # splines or hierarchy), so this can be done blindly
-    opts = dict(default_blend_opts)
+    opts = {}
     for d in src, dst, edit:
         opts.update(d.get('blend', {}))
+    opts = Wrapper(opts, specs.blend)
 
-    blended = merge_nodes(specs.node, src, dst, edit, opts['nloops'])
-    name_map = sort_xforms(src['xforms'], dst['xforms'], opts['xform_sort'],
-                           explicit=zip(*opts.get('xform_map', [])))
+    blended = merge_nodes(specs.node, src, dst, edit, opts.nloops)
+    name_map = sort_xforms(src['xforms'], dst['xforms'], opts.xform_sort,
+                           explicit=zip(*opts.xform_map))
 
     blended['xforms'] = {}
     for (sxf_key, dxf_key) in name_map:
@@ -49,7 +111,7 @@ def blend(src, dst, edit={}):
         blended['xforms'][bxf_key] = blend_xform(
                 src['xforms'].get(sxf_key),
                 dst['xforms'].get(dxf_key),
-                xf_edits, opts['nloops'])
+                xf_edits, opts.nloops)
 
     if 'final_xform' in src or 'final_xform' in dst:
         blended['final_xform'] = blend_xform(src.get('final_xform'),
@@ -58,7 +120,7 @@ def blend(src, dst, edit={}):
     # TODO: write 'info' section
     # TODO: palflip
     blended['type'] = 'animation'
-    blended.setdefault('time', {})['duration'] = opts['duration']
+    blended.setdefault('time', {})['duration'] = opts.duration
     return blended
 
 def merge_edits(sv, av, bv):
@@ -75,16 +137,16 @@ def merge_edits(sv, av, bv):
     else:
         return bv if bv is not None else av
 
-def tospline(spl, src, dst, edit, loops):
-    def split_node_val(val):
-        if val is None:
-            return spl.default, 0
-        if isinstance(val, (int, float)):
-            return val, 0
-        return val
+def split_node_val(spl, val):
+    if val is None:
+        return spl.default, 0
+    if isinstance(val, (int, float)):
+        return val, 0
+    return val
 
-    sp, sv = split_node_val(src)    # position, velocity
-    dp, dv = split_node_val(dst)
+def tospline(spl, src, dst, edit, loops):
+    sp, sv = split_node_val(spl, src)    # position, velocity
+    dp, dv = split_node_val(spl, dst)
 
     # For variation parameters, copy missing values instead of using defaults
     if spl.var:
@@ -285,4 +347,4 @@ def palflip(gnm):
 if __name__ == "__main__":
     import sys, json
     a, b, c = [json.load(open(f+'.json')) for f in 'abc']
-    print util.json_encode(blend(a, b, c))
+    print json_encode(blend(a, b, c))
diff --git a/cuburn/genome/convert.py b/cuburn/genome/convert.py
index ecde73f..b66ddd8 100644
--- a/cuburn/genome/convert.py
+++ b/cuburn/genome/convert.py
@@ -8,6 +8,9 @@ import numpy as np
 from variations import var_params
 import util
 
+# Re-exported
+from blend import node_to_anim, edge_to_anim
+
 class XMLGenomeParser(object):
     """
     Parse an XML genome into a list of dictionaries.
@@ -163,8 +166,10 @@ def apply_structure(struct, src):
             out[l[0]] = v
     return out
 
-def convert_flame(flame):
-    return util.unflatten(util.flatten(apply_structure(flame_structure, flame)))
+def flam3_to_node(flame):
+    n = util.unflatten(util.flatten(apply_structure(flame_structure, flame)))
+    n['type'] = 'node'
+    return n
 
 def convert_file(path):
     """Quick one-shot conversion for an XML genome."""
@@ -173,7 +178,7 @@ def convert_file(path):
         warnings.warn("Lot of flames in this file. Sure it's not a "
                       "frame-based animation?")
     for flame in flames:
-        yield convert_flame(flame)
+        yield flam3_to_node(flame)
 
 if __name__ == "__main__":
     import sys
diff --git a/cuburn/genome/specs.py b/cuburn/genome/specs.py
index fb136e8..9f420c4 100644
--- a/cuburn/genome/specs.py
+++ b/cuburn/genome/specs.py
@@ -68,6 +68,13 @@ time = (
   , 'frame_width': scalespline(d='Scale of profile temporal width per frame.')
   })
 
+blend = (
+  { 'nloops': scalar(2)
+  , 'duration': scalar(2)
+  , 'xform_sort': enum('weightflip weight natural color')
+  , 'xform_map': list_(list_(String('xfid'), d='A pair of src, dst IDs'))
+  })
+
 base = (
   { 'name': String("Human-readable name of this work")
   , 'camera': camera
diff --git a/cuburn/genome/use.py b/cuburn/genome/use.py
index bc45a5c..b5ebd46 100644
--- a/cuburn/genome/use.py
+++ b/cuburn/genome/use.py
@@ -62,6 +62,19 @@ class Wrapper(object):
             return self.spec.type
         return self.spec[name]
 
+    @classmethod
+    def visit(cls, obj):
+        """
+        Visit every node. Note that for simplicity, this function will be
+        called on all elements (i.e. pivoting to a new Wrapper type inside the
+        wrapping function and overriding visit() won't do anything).
+        """
+        if isinstance(obj, (Wrapper, dict)):
+            return dict((k, cls.visit(obj[k])) for k in obj)
+        elif isinstance(obj, list):
+            return [cls.visit(o) for o in obj]
+        return obj
+
     def __getattr__(self, name):
         return self.wrap(name, self.get_spec(name), self._val.get(name))
 
@@ -155,6 +168,7 @@ class SplineEval(object):
         return times, vals, t, scale
 
     def __call__(self, itime, deriv=0):
+        # TODO: respect 'interp' THIS IS IMPORTANT.
         times, vals, t, scale = self.find_knots(itime)
 
         m1 = (vals[2] - vals[0]) / (1.0 - times[0])
diff --git a/cuburn/genome/util.py b/cuburn/genome/util.py
index 32e0ddd..51c193c 100644
--- a/cuburn/genome/util.py
+++ b/cuburn/genome/util.py
@@ -2,6 +2,7 @@ import base64
 import numpy as np
 
 from cuburn.code.util import crep
+import spectypes
 
 def get(dct, default, *keys):
     if len(keys) == 1:
@@ -49,6 +50,13 @@ def unflatten(kvlist):
         go(out, k.split('.'), v)
     return out
 
+def resolve_spec(sp, path):
+    for name in path:
+        if isinstance(sp, spectypes.Map):
+            sp = sp.type
+        else:
+            sp = sp[name]
+    return sp
 
 def palette_decode(datastrs):
     """