diff --git a/comfy/model_management.py b/comfy/model_management.py
index 64ed197..53582fc 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -339,7 +339,11 @@ def free_memory(memory_required, device, keep_loaded=[]):
 
     if unloaded_model:
         soft_empty_cache()
-
+    else:
+        if vram_state != VRAMState.HIGH_VRAM:
+            mem_free_total, mem_free_torch = get_free_memory(device, torch_free_too=True)
+            if mem_free_torch > mem_free_total * 0.25:
+                soft_empty_cache()
 
 def load_models_gpu(models, memory_required=0):
     global vram_state