diff --git a/modules/intel/openvino/__init__.py b/modules/intel/openvino/__init__.py
index 5023a2dcd..c2af24637 100644
--- a/modules/intel/openvino/__init__.py
+++ b/modules/intel/openvino/__init__.py
@@ -450,7 +450,7 @@ def openvino_fx(subgraph, example_inputs):
                                 example_inputs_reordered.append(example_inputs[idx1])
                 example_inputs = example_inputs_reordered
 
-            if dont_use_faketensors:
+            if dont_use_faketensors or shared.opts.openvino_disable_memory_cleanup:
                 pass
             else:
                 # Delete unused subgraphs
diff --git a/modules/processing_diffusers.py b/modules/processing_diffusers.py
index fdf29f082..96183f3f3 100644
--- a/modules/processing_diffusers.py
+++ b/modules/processing_diffusers.py
@@ -317,12 +317,12 @@ def process_diffusers(p: processing.StableDiffusionProcessing):
         if shared.opts.cuda_compile and shared.opts.cuda_compile_backend == "openvino_fx":
             if shared.compiled_model_state.first_pass and op == "base":
                 shared.compiled_model_state.first_pass = False
-                if hasattr(shared.sd_model, "unet"):
+                if not shared.opts.openvino_disable_memory_cleanup and hasattr(shared.sd_model, "unet"):
                     shared.sd_model.unet.apply(sd_models.convert_to_faketensors)
                     devices.torch_gc(force=True)
             if shared.compiled_model_state.first_pass_refiner and op == "refiner":
                 shared.compiled_model_state.first_pass_refiner = False
-                if hasattr(shared.sd_refiner, "unet"):
+                if not shared.opts.openvino_disable_memory_cleanup and hasattr(shared.sd_refiner, "unet"):
                     shared.sd_refiner.unet.apply(sd_models.convert_to_faketensors)
                     devices.torch_gc(force=True)
 
diff --git a/modules/processing_vae.py b/modules/processing_vae.py
index 55c545cad..985bcdac4 100644
--- a/modules/processing_vae.py
+++ b/modules/processing_vae.py
@@ -51,7 +51,7 @@ def full_vae_decode(latents, model):
     # Delete PyTorch VAE after OpenVINO compile
     if shared.opts.cuda_compile and shared.opts.cuda_compile_backend == "openvino_fx" and shared.compiled_model_state.first_pass_vae:
         shared.compiled_model_state.first_pass_vae = False
-        if hasattr(shared.sd_model, "vae"):
+        if not shared.opts.openvino_disable_memory_cleanup and hasattr(shared.sd_model, "vae"):
             model.vae.apply(sd_models.convert_to_faketensors)
             devices.torch_gc(force=True)
 
diff --git a/modules/shared.py b/modules/shared.py
index 7082be669..c986333cb 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -390,6 +390,7 @@ options_templates.update(options_section(('cuda', "Compute Settings"), {
     "nncf_compress_weights_mode": OptionInfo("INT8", "OpenVINO compress mode for NNCF", gr.Radio, {"choices": ['INT8', 'INT8_SYM', 'INT4_ASYM', 'INT4_SYM', 'NF4'], "visible": cmd_opts.use_openvino}),
     "nncf_compress_weights_raito": OptionInfo(1.0, "OpenVINO compress ratio for NNCF", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.01, "visible": cmd_opts.use_openvino}),
     "openvino_disable_model_caching": OptionInfo(False, "OpenVINO disable model caching", gr.Checkbox, {"visible": cmd_opts.use_openvino}),
+    "openvino_disable_memory_cleanup": OptionInfo(True, "OpenVINO disable memory cleanup after compile", gr.Checkbox, {"visible": cmd_opts.use_openvino}),
 
     "directml_sep": OptionInfo("<h2>DirectML</h2>", "", gr.HTML, {"visible": devices.backend == "directml"}),
     "directml_memory_provider": OptionInfo(default_memory_provider, 'DirectML memory stats provider', gr.Radio, {"choices": memory_providers, "visible": devices.backend == "directml"}),