diff --git a/gpu/gpu.go b/gpu/gpu.go
index 64395eaf..db914552 100644
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -869,6 +869,13 @@ func splitTransform(t f32.Affine2D) (srs f32.Affine2D, offset f32.Point) {
 	return
 }
 
+func (d *drawOps) save(id int, state drawState) {
+	if extra := id - len(d.states) + 1; extra > 0 {
+		d.states = append(d.states, make([]drawState, extra)...)
+	}
+	d.states[id] = state
+}
+
 func (d *drawOps) collectOps(r *ops.Reader, state drawState) {
 	var (
 		quads  quadsOp
@@ -876,6 +883,7 @@ func (d *drawOps) collectOps(r *ops.Reader, state drawState) {
 		dashes dashOp
 		z      int
 	)
+	d.save(opconst.InitialStateID, state)
 loop:
 	for encOp, ok := r.Decode(); ok; encOp, ok = r.Decode() {
 		switch opconst.OpType(encOp.Data[0]) {
@@ -1029,13 +1037,16 @@ loop:
 			}
 		case opconst.TypeSave:
 			id := ops.DecodeSave(encOp.Data)
-			if extra := id - len(d.states) + 1; extra > 0 {
-				d.states = append(d.states, make([]drawState, extra)...)
-			}
-			d.states[id] = state
+			d.save(id, state)
 		case opconst.TypeLoad:
-			id := ops.DecodeLoad(encOp.Data)
-			state = d.states[id]
+			id, mask := ops.DecodeLoad(encOp.Data)
+			s := d.states[id]
+			if mask&opconst.TransformState != 0 {
+				state.t = s.t
+			}
+			if mask&^opconst.TransformState != 0 {
+				state = s
+			}
 		}
 	}
 }
diff --git a/internal/opconst/ops.go b/internal/opconst/ops.go
index cff4b0d1..58904ee1 100644
--- a/internal/opconst/ops.go
+++ b/internal/opconst/ops.go
@@ -57,7 +57,7 @@ const (
 	TypeKeyFocusLen        = 1 + 1
 	TypeKeySoftKeyboardLen = 1 + 1
 	TypeSaveLen            = 1 + 4
-	TypeLoadLen            = 1 + 4
+	TypeLoadLen            = 1 + 1 + 4
 	TypeAuxLen             = 1
 	TypeClipLen            = 1 + 4*4 + 1
 	TypeProfileLen         = 1
@@ -67,6 +67,20 @@ const (
 	TypeDashLen            = 1 + 4 + 1
 )
 
+// StateMask is a bitmask of state types a load operation
+// should restore.
+type StateMask uint8
+
+const (
+	TransformState StateMask = 1 << iota
+
+	AllState = ^StateMask(0)
+)
+
+// InitialStateID is the ID for saving and loading
+// the initial operation state.
+const InitialStateID = 0
+
 func (t OpType) Size() int {
 	return [...]int{
 		TypeMacroLen,
diff --git a/internal/ops/ops.go b/internal/ops/ops.go
index aa5b571c..49f68d73 100644
--- a/internal/ops/ops.go
+++ b/internal/ops/ops.go
@@ -72,11 +72,11 @@ func DecodeSave(data []byte) int {
 	return int(bo.Uint32(data[1:]))
 }
 
-// DecodeLoad decodes the state id of a restore op.
-func DecodeLoad(data []byte) int {
+// DecodeLoad decodes the state id and mask of a load op.
+func DecodeLoad(data []byte) (int, opconst.StateMask) {
 	if opconst.OpType(data[0]) != opconst.TypeLoad {
 		panic("invalid op")
 	}
 	bo := binary.LittleEndian
-	return int(bo.Uint32(data[1:]))
+	return int(bo.Uint32(data[2:])), opconst.StateMask(data[1])
 }
diff --git a/internal/rendertest/refs/TestDeferredPaint.png b/internal/rendertest/refs/TestDeferredPaint.png
index 68f278d5..1446810f 100644
Binary files a/internal/rendertest/refs/TestDeferredPaint.png and b/internal/rendertest/refs/TestDeferredPaint.png differ
diff --git a/internal/rendertest/render_test.go b/internal/rendertest/render_test.go
index 33c1ff43..9cdb68ec 100644
--- a/internal/rendertest/render_test.go
+++ b/internal/rendertest/render_test.go
@@ -107,13 +107,17 @@ func TestDeferredPaint(t *testing.T) {
 	run(t, func(o *op.Ops) {
 		state := op.Save(o)
 		clip.Rect(image.Rect(0, 0, 80, 80)).Op().Add(o)
-		paint.ColorOp{Color: color.NRGBA{A: 0xff, R: 0xff}}.Add(o)
+		paint.ColorOp{Color: color.NRGBA{A: 0xff, G: 0xff}}.Add(o)
+		paint.PaintOp{}.Add(o)
+
+		op.Affine(f32.Affine2D{}.Offset(f32.Pt(20, 20))).Add(o)
 		m := op.Record(o)
+		clip.Rect(image.Rect(0, 0, 80, 80)).Op().Add(o)
+		paint.ColorOp{Color: color.NRGBA{A: 0xff, R: 0xff, G: 0xff}}.Add(o)
 		paint.PaintOp{}.Add(o)
 		paintMacro := m.Stop()
 		op.Defer(o, paintMacro)
-		paint.ColorOp{Color: color.NRGBA{A: 0xff, G: 0xff}}.Add(o)
-		paint.PaintOp{}.Add(o)
+
 		state.Load()
 		op.Affine(f32.Affine2D{}.Offset(f32.Pt(10, 10))).Add(o)
 		clip.Rect(image.Rect(0, 0, 80, 80)).Op().Add(o)
diff --git a/io/router/key.go b/io/router/key.go
index cb50d079..5f64824c 100644
--- a/io/router/key.go
+++ b/io/router/key.go
@@ -140,7 +140,7 @@ func (q *keyQueue) resolveFocus(events *handlerEvents) resolveState {
 			q.states[id] = state
 			state = resolveState{}
 		case opconst.TypeLoad:
-			id := ops.DecodeLoad(encOp.Data)
+			id, mask := ops.DecodeLoad(encOp.Data)
 			restored := q.states[id]
 			if state.keyboard > restored.keyboard {
 				restored.keyboard = state.keyboard
@@ -148,7 +148,9 @@ func (q *keyQueue) resolveFocus(events *handlerEvents) resolveState {
 			if state.pri.replaces(restored.pri) {
 				restored.tag, restored.pri = state.tag, state.pri
 			}
-			state = restored
+			if mask != 0 {
+				state = restored
+			}
 		}
 	}
 	return state
diff --git a/io/router/pointer.go b/io/router/pointer.go
index f0115b66..0fb2b22e 100644
--- a/io/router/pointer.go
+++ b/io/router/pointer.go
@@ -85,22 +85,33 @@ const (
 	areaEllipse
 )
 
+func (q *pointerQueue) save(id int, state collectState) {
+	if extra := id - len(q.states) + 1; extra > 0 {
+		q.states = append(q.states, make([]collectState, extra)...)
+	}
+	q.states[id] = state
+}
+
 func (q *pointerQueue) collectHandlers(r *ops.Reader, events *handlerEvents) {
 	state := collectState{
 		area: -1,
 		node: -1,
 	}
+	q.save(opconst.InitialStateID, state)
 	for encOp, ok := r.Decode(); ok; encOp, ok = r.Decode() {
 		switch opconst.OpType(encOp.Data[0]) {
 		case opconst.TypeSave:
 			id := ops.DecodeSave(encOp.Data)
-			if extra := id - len(q.states) + 1; extra > 0 {
-				q.states = append(q.states, make([]collectState, extra)...)
-			}
-			q.states[id] = state
+			q.save(id, state)
 		case opconst.TypeLoad:
-			id := ops.DecodeLoad(encOp.Data)
-			state = q.states[id]
+			id, mask := ops.DecodeLoad(encOp.Data)
+			s := q.states[id]
+			if mask&opconst.TransformState != 0 {
+				state.t = s.t
+			}
+			if mask&^opconst.TransformState != 0 {
+				state = s
+			}
 		case opconst.TypePass:
 			state.pass = encOp.Data[1] != 0
 		case opconst.TypeArea:
diff --git a/op/op.go b/op/op.go
index efe8cbc2..7dc0af02 100644
--- a/op/op.go
+++ b/op/op.go
@@ -144,8 +144,8 @@ type pc struct {
 
 // Defer executes c after all other operations have completed,
 // including previously deferred operations.
-// Defer saves the operation state, which is then loaded prior
-// to execution.
+// Defer saves the current transformation and restores it prior
+// to execution. All other operation state is reset.
 //
 // Note that deferred operations are executed in first-in-first-out
 // order, unlike the Go facility of the same name.
@@ -156,8 +156,8 @@ func Defer(o *Ops, c CallOp) {
 	state := Save(o)
 	// Wrap c in a macro that loads the saved state before execution.
 	m := Record(o)
-	//
-	state.load()
+	load(o, opconst.InitialStateID, opconst.AllState)
+	load(o, state.id, opconst.TransformState)
 	c.Add(o)
 	c = m.Stop()
 	// A Defer is recorded as a TypeDefer followed by the
@@ -175,11 +175,17 @@ func Save(o *Ops) StateOp {
 		id:      o.nextStateID,
 		macroID: o.macroStack.currentID,
 	}
+	save(o, s.id)
+	return s
+}
+
+// save records a save of the operations state to
+// id.
+func save(o *Ops, id int) {
 	bo := binary.LittleEndian
 	data := o.Write(opconst.TypeSaveLen)
 	data[0] = byte(opconst.TypeSave)
-	bo.PutUint32(data[1:], uint32(s.id))
-	return s
+	bo.PutUint32(data[1:], uint32(id))
 }
 
 // Load a previously saved operations state.
@@ -187,18 +193,20 @@ func (s StateOp) Load() {
 	if s.ops.macroStack.currentID != s.macroID {
 		panic("load in a different macro than save")
 	}
-	s.load()
-}
-
-// load is like Load without the same-macro check.
-func (s StateOp) load() {
 	if s.id == 0 {
 		panic("zero-value op")
 	}
+	load(s.ops, s.id, opconst.AllState)
+}
+
+// load a previously saved operations state given
+// its ID. Only state included in mask is affected.
+func load(o *Ops, id int, m opconst.StateMask) {
 	bo := binary.LittleEndian
-	data := s.ops.Write(opconst.TypeLoadLen)
+	data := o.Write(opconst.TypeLoadLen)
 	data[0] = byte(opconst.TypeLoad)
-	bo.PutUint32(data[1:], uint32(s.id))
+	data[1] = byte(m)
+	bo.PutUint32(data[2:], uint32(id))
 }
 
 // Reset the Ops, preparing it for re-use. Reset invalidates