diff --git a/comfy/model_management.py b/comfy/model_management.py
index d5d77d7d737..6f667dfc57e 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -314,6 +314,9 @@ def model(self):
     def model_memory(self):
         return self.model.model_size()
 
+    def model_loaded_memory(self):
+        return self.model.loaded_size()
+
     def model_offloaded_memory(self):
         return self.model.model_size() - self.model.loaded_size()
 
@@ -504,8 +507,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         lowvram_model_memory = 0
         if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM) and not force_full_load:
             model_size = loaded_model.model_memory_required(torch_dev)
-            current_free_mem = get_free_memory(torch_dev)
-            lowvram_model_memory = max(1, (current_free_mem - minimum_memory_required), min(current_free_mem * 0.4, current_free_mem - minimum_inference_memory()))
+            loaded_memory = loaded_model.model_loaded_memory()
+            current_free_mem = get_free_memory(torch_dev) + loaded_memory
+            lowvram_model_memory = max(64 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * 0.4, current_free_mem - minimum_inference_memory()))
+            lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory)
             if model_size <= lowvram_model_memory: #only switch to lowvram if really necessary
                 lowvram_model_memory = 0