fix

app.py

@@ -1,7 +1,8 @@
+import spaces
 import torch
 from pipelines.inverted_ve_pipeline import STYLE_DESCRIPTION_DICT, create_image_grid
 import gradio as gr
-import os, json, cv2
+import os, json
 import numpy as np
 from PIL import Image
 
@@ -107,12 +108,13 @@ def load_example_controlnet():
             inf_object_name = config["inference_info"]["inf_object_list"][0]
 
             canny_path = './assets/depth_dir/gundam.png'
-            image_info = [image_path, canny_path, style_name, "", 1, 0.5, 50] # empty text
+            image_info = [image_path, canny_path, style_name, "", 1, 0.5, 50]
 
             examples.append(image_info)
 
     return examples
 
+@spaces.GPU
 def controlnet_fn(image_path, depth_image_path, style_name, content_text, output_number, controlnet_scale=0.5, diffusion_step=50):
     """
 
@@ -215,7 +217,6 @@ def controlnet_fn(image_path, depth_image_path, style_name, content_text, output
                 # make grid
                 grid = create_image_grid(images, n_row, n_col)
 
-        torch.cuda.empty_cache()
         return grid
 
 
@@ -225,16 +226,16 @@ description_md = """
 ### 📖 [[Paper](https://arxiv.org/abs/2402.12974)] | ✨ [[Project page](https://curryjung.github.io/VisualStylePrompt)] | ✨ [[Code](https://github.com/naver-ai/Visual-Style-Prompting)]
 ### 🔥 [[Default ver](https://huggingface.co/spaces/naver-ai/VisualStylePrompting)]
 ---
-### Visual Style Prompting also works on `ControlNet` which specifies the shape of the results by depthmap or keypoints.
-
-### To try out our demo with ControlNet,
+### ✨ Visual Style Prompting also works on `ControlNet` which specifies the shape of the results by depthmap or keypoints.
+### ‼️ w/ ControlNet ver does not support user style images.
+### 🔥 To try out our demo with ControlNet,
 1. Upload an `image for depth control`. An off-the-shelf model will produce the depthmap from it.
 2. Choose `ControlNet scale` which determines the alignment to the depthmap.
 3. Choose a `style reference` from the collection of images below.
 4. Enter the `text prompt`. (`Empty text` is okay, but a depthmap description helps.)
 5. Choose the `number of outputs`.
 
-### To achieve faster results, we recommend lowering the diffusion steps to 30.
+### 👉 To achieve faster results, we recommend lowering the diffusion steps to 30.
 ### Enjoy ! 😄
 """
 
@@ -249,7 +250,7 @@ iface_controlnet = gr.Interface(
         gr.components.Slider(minimum=0.5, maximum=10, step=0.5, value=0.5, label="Controlnet scale"),
         gr.components.Slider(minimum=10, maximum=50, step=10, value=50, label="Diffusion steps")
     ],
-    outputs=gr.components.Image(type="pil"),
+    outputs=gr.components.Image(label="Generated Image"),
     title="🎨 Visual Style Prompting (w/ ControlNet)",
     description=description_md,
     examples=load_example_controlnet(),

requirements.txt

@@ -1,4 +1,4 @@
-torch==1.13.1
+torch
 diffusers
 transformers
 accelerate
@@ -8,5 +8,5 @@ gradio
 triton
 torchvision
 opencv-python
-xformers==0.0.16
+xformers