diff --git a/src/aspire/commands/cov3d.py b/src/aspire/commands/cov3d.py
index 3a0752740..61fca0696 100644
--- a/src/aspire/commands/cov3d.py
+++ b/src/aspire/commands/cov3d.py
@@ -52,10 +52,10 @@ def cov3d(
 
     source = source.whiten()
     basis = FBBasis3D((max_resolution, max_resolution, max_resolution))
-    mean_estimator = MeanEstimator(source, basis, batch_size=8192)
+    mean_estimator = MeanEstimator(source, basis, batch_size=512)
     mean_est = mean_estimator.estimate()
 
-    noise_estimator = WhiteNoiseEstimator(source, batch_size=500)
+    noise_estimator = WhiteNoiseEstimator(source, batch_size=512)
     # Estimate the noise variance. This is needed for the covariance estimation step below.
     noise_variance = noise_estimator.estimate()
     logger.info(f"Noise Variance = {noise_variance}")
diff --git a/src/aspire/covariance/covar2d.py b/src/aspire/covariance/covar2d.py
index 30106909a..d0bb73b57 100644
--- a/src/aspire/covariance/covar2d.py
+++ b/src/aspire/covariance/covar2d.py
@@ -513,10 +513,14 @@ class BatchedRotCov2D(RotCov2D):
         be extracted.
     :param basis: The `FBBasis2D` object used to decompose the images. By
         default, this is set to `FFBBasis2D((src.L, src.L))`.
-    :param batch_size: The number of images to process at a time (default 8192).
+    :param batch_size: The number of images to process at a time (default 512).
+        512 is a good starting point for large images with a GPU where
+        memory is a concern.  If the GPU runs out of memory, try
+        scaling down `batch_size`.  For hi-memory CPU applications,
+        scaling up to a larger value such as 8192 may yield better performance.
     """
 
-    def __init__(self, src, basis=None, batch_size=8192):
+    def __init__(self, src, basis=None, batch_size=512):
         self.src = src
         self.basis = basis
         self.batch_size = batch_size