diff --git a/CHANGELOG.md b/CHANGELOG.md
index a3fb7ab6..f9dbdead 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,8 @@
+## 1.5.3
+
+### Enhancement
+- Free intermediate arrays (`origin_img`, `img`, `ort_inputs`, `output`) at dead points during YoloX `image_processing()` to reduce peak memory during inference
+
 ## 1.5.2
 
 ### Fix
diff --git a/unstructured_inference/__version__.py b/unstructured_inference/__version__.py
index 928be505..e375e78d 100644
--- a/unstructured_inference/__version__.py
+++ b/unstructured_inference/__version__.py
@@ -1 +1 @@
-__version__ = "1.5.2"  # pragma: no cover
+__version__ = "1.5.3"  # pragma: no cover
diff --git a/unstructured_inference/models/yolox.py b/unstructured_inference/models/yolox.py
index 932242ec..d53bc085 100644
--- a/unstructured_inference/models/yolox.py
+++ b/unstructured_inference/models/yolox.py
@@ -108,12 +108,15 @@ def image_processing(
         input_shape = (1024, 768)
         origin_img = np.array(image)
         img, ratio = preprocess(origin_img, input_shape)
+        del origin_img  # Free full-size image array before ONNX inference
         session = self.model
 
         ort_inputs = {session.get_inputs()[0].name: img[None, :, :, :]}
         output = session.run(None, ort_inputs)
+        del img, ort_inputs  # Free preprocessed inputs after inference
         # TODO(benjamin): check for p6
         predictions = demo_postprocess(output[0], input_shape, p6=False)[0]
+        del output
 
         boxes = predictions[:, :4]
         scores = predictions[:, 4:5] * predictions[:, 5:]