diff --git a/src/aspire/basis/ffb_2d.py b/src/aspire/basis/ffb_2d.py
index 727a1e0b7..77c48608b 100644
--- a/src/aspire/basis/ffb_2d.py
+++ b/src/aspire/basis/ffb_2d.py
@@ -280,16 +280,28 @@ def _filter_stack_to_basis_mats(self, f, **kwargs):
         omegay = k * xp.sin(theta)
         omega = 2 * np.pi * xp.vstack((omegax.flatten("C"), omegay.flatten("C")))
 
+        # For high non-radial filter counts at higher pixel counts
+        # h_vals2d requires a large amount of memory and is too large
+        # to fit on a GPU
+        # In the smaller cases, the code attepts using GPU.
+        if len(f) * omega.size >= self.MAX_GPU_ELEM_COUNT:
+            # when too large put omega on host
+            omega = xp.asnumpy(omega)
+
         # This should return either a single 2d array, or stack of 2d arrays
-        # Reshape singleton to stack of 1.
-        h_vals2d = (
-            h_fun(omega, pixel_size=pixel_size)
-            .reshape(len(f), n_k, n_theta)
-            .astype(self.dtype)
-        )
+        # Code will reshape singleton to stack of 1 below.
+        h_vals2d = h_fun(omega, pixel_size=pixel_size)
+
+        # when possible, perform sum via GPU
+        if len(f) * omega.size < self.MAX_GPU_ELEM_COUNT:
+            h_vals2d = xp.asarray(h_vals2d)
+
+        h_vals2d = h_vals2d.reshape(len(f), n_k, n_theta).astype(self.dtype, copy=False)
         h_vals = h_vals2d.sum(axis=-1) / n_theta
+        h_vals = xp.asarray(h_vals)  # no-op if already fit on GPU
 
         h_basis = self.expand_radial_vec(h_vals, **kwargs)
+
         return h_basis
 
     def filter_to_basis_mat(self, f, **kwargs):
diff --git a/src/aspire/basis/fle_2d.py b/src/aspire/basis/fle_2d.py
index d91fea604..039160d88 100644
--- a/src/aspire/basis/fle_2d.py
+++ b/src/aspire/basis/fle_2d.py
@@ -17,9 +17,6 @@
 
 logger = logging.getLogger(__name__)
 
-# Number of elements in filter_to_basis_mat before breaking into batches
-MAX_ELEM_COUNT = 2e9
-
 
 def _cleanup():
     """
@@ -848,14 +845,14 @@ def _filter_stack_to_basis_mats(self, f, **kwargs):
         # h_vals2d requires a large amount of memory and is too large
         # to fit on a GPU
         # In the smaller cases, the code attepts using GPU.
-        if len(f) * omega.size >= MAX_ELEM_COUNT:
+        if len(f) * omega.size >= self.MAX_GPU_ELEM_COUNT:
             # when too large put omega on host
             omega = xp.asnumpy(omega)
 
         h_vals2d = h_fun(omega, pixel_size=pixel_size)
 
         # when possible, perform sum via GPU
-        if len(f) * omega.size < MAX_ELEM_COUNT:
+        if len(f) * omega.size < self.MAX_GPU_ELEM_COUNT:
             h_vals2d = xp.asarray(h_vals2d)
 
         h_vals2d = h_vals2d.reshape(len(f), n_k, n_theta).astype(self.dtype, copy=False)
diff --git a/src/aspire/basis/steerable.py b/src/aspire/basis/steerable.py
index 8429b3157..cae7fe6db 100644
--- a/src/aspire/basis/steerable.py
+++ b/src/aspire/basis/steerable.py
@@ -21,6 +21,9 @@ class SteerableBasis2D(Basis, abc.ABC):
     # Default matrix type for basis representation.
     matrix_type = BlkDiagMatrix
 
+    # Number of elements in filter_to_basis_mat before breaking into batches
+    MAX_GPU_ELEM_COUNT = 2e9
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)