diff --git a/tinygrad/codegen/__init__.py b/tinygrad/codegen/__init__.py
index 0d6f4cc670..c823ef4777 100644
--- a/tinygrad/codegen/__init__.py
+++ b/tinygrad/codegen/__init__.py
@@ -57,6 +57,11 @@ pm_no_weakints = PatternMatcher([
   (UPat(GroupOp.All, dtype=dtypes.weakint, name="x"), lambda x: x.replace(dtype=dtypes.int))
 ])
 
+pm_fix_image_shrink = PatternMatcher([
+  (UPat(Ops.SHRINK, src=(UPat(Ops.RESHAPE, src=(UPat(Ops.PARAM, name="img"), UPat())), UPat(name="idx"), UPat()), name="out"),
+   lambda img,out,idx: img.reshape(-1, 4).index(idx.src[0]) if len(img.shape) == 3 else None),
+])
+
 def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   if VIZ: graph_rewrite(ast, PatternMatcher([]), name="View Base AST")
   if DEBUG >= 5: print(pyrender(ast))
@@ -103,10 +108,6 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   # add loads and remove invalids
   sink = graph_rewrite(sink, pm_add_loads+pm_remove_invalid, name="** add loads (code)")
 
-  # create image buffers
-  if IMAGE and ren.target.device in {"QCOM", "CL", "PYTHON", "NULL"}:
-    sink = graph_rewrite(sink, pm_make_images, name="create image buffers", bottom_up=True, ctx=ren.target.arch)
-
   # devectorize
   sink = graph_rewrite(sink, sym+devectorize_alu+devectorize_buf_and_index+load_store_folding+correct_load_store+load_store_indexing,
                        ctx=ren, name="devectorize")
@@ -124,9 +125,16 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   sink = graph_rewrite(sink, pm_decomp, name="early decompositions")
   sink = graph_rewrite(sink, pm_dtype_decomps, ctx=(set(), ren), name="decomp dtypes")
 
+  # create image buffers
+  if IMAGE and ren.target.device in {"QCOM", "CL", "PYTHON", "NULL"}:
+    sink = graph_rewrite(sink, pm_make_images, name="create image buffers", bottom_up=True, ctx=ren.target.arch)
+
   # do memory coalesing (late)
   sink = memory_coalesing(sink, ren)
 
+  # image fixup
+  sink = graph_rewrite(sink, pm_mops+pm_fix_image_shrink, name="fix image shrink")
+
   # instruction selection decompositions
   pm_decomp = pm_decomp+\
     get_late_rewrite_patterns(supported_ops, bool(DISABLE_FAST_IDIV))+\
diff --git a/tinygrad/codegen/late/coalese.py b/tinygrad/codegen/late/coalese.py
index fb939dca29..78494a76a0 100644
--- a/tinygrad/codegen/late/coalese.py
+++ b/tinygrad/codegen/late/coalese.py
@@ -13,7 +13,7 @@ def memory_coalesing(sink:UOp, ctx:Renderer) -> UOp:
   memory: defaultdict[tuple[Ops, UOp, Any, Any], dict[int, list[UOp]]]  = defaultdict(dict)
   for u in sink.toposort():
     # TODO: this should handle images too, it's just memory coalesing
-    if u.op in {Ops.LOAD, Ops.STORE} and not isinstance(u.src[0].src[0].dtype, ImageDType):
+    if u.op in {Ops.LOAD, Ops.STORE}:
       assert len(u.src) == (2 if u.op is Ops.STORE else 1), "memory coalesing does not support gated loads/stores"
       if u.src[0].op is not Ops.INDEX: continue
       buf, idx_u = u.src[0].src