diff --git a/api/onnx_web/convert/diffusion/textual_inversion.py b/api/onnx_web/convert/diffusion/textual_inversion.py
index cef2206e..8ba3e488 100644
--- a/api/onnx_web/convert/diffusion/textual_inversion.py
+++ b/api/onnx_web/convert/diffusion/textual_inversion.py
@@ -74,6 +74,8 @@ def convert_diffusion_textual_inversion(
         return_tensors="pt",
     )
 
+    tokenizer.save_pretrained(path.join(dest_path, "tokenizer"))
+
     export(
         text_encoder,
         # casting to torch.int32 until the CLIP fix is released: https://github.com/huggingface/transformers/pull/18515/files
diff --git a/api/onnx_web/diffusion/load.py b/api/onnx_web/diffusion/load.py
index e3b54510..1845c50a 100644
--- a/api/onnx_web/diffusion/load.py
+++ b/api/onnx_web/diffusion/load.py
@@ -21,6 +21,7 @@ from diffusers import (
     PNDMScheduler,
     StableDiffusionPipeline,
 )
+from transformers import CLIPTokenizer
 
 try:
     from diffusers import DEISMultistepScheduler
@@ -200,6 +201,9 @@ def load_pipeline(
                 provider=device.ort_provider(),
                 sess_options=device.sess_options(),
             )
+            components["tokenizer"] = CLIPTokenizer.from_pretrained(
+                path.join(inversion, "tokenizer"),
+            )
 
         pipe = pipeline.from_pretrained(
             model,