<fix> fix some bugs.

app.py

@@ -47,7 +47,9 @@ The corresponding condition images will be automatically extracted.
 
 
 def init_basemodel():
-    global transformer, scheduler, vae, text_encoder, text_encoder_2, tokenizer, tokenizer_2, image_processor
+    global transformer, scheduler, vae, text_encoder, text_encoder_2, tokenizer, tokenizer_2, image_processor, pipe, current_task
+    pipe = None
+    current_task = None
 
     # init models
     scheduler = diffusers.FlowMatchEulerDiscreteScheduler()
@@ -77,8 +79,12 @@ def init_basemodel():
 @spaces.GPU
 def process_image_and_text(condition_image, target_prompt, condition_image_prompt, task, random_seed, inpainting, fill_x1, fill_x2, fill_y1, fill_y2):
     # set up the model
+    global pipe, current_task
     if pipe is None or current_task != task:
+        current_task = task
+
         # init transformer
+        global transformer
         transformer = HunyuanVideoTransformer3DModel.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
                                                                      subfolder="transformer", 
                                                                      inference_subject_driven=task in ['subject_driven'])
@@ -181,7 +187,6 @@ def process_image_and_text(condition_image, target_prompt, condition_image_promp
                 img_gray = cv2.cvtColor(img_np, cv2.COLOR_BGR2GRAY)
                 edges = cv2.Canny(img_gray, 100, 200)
                 edges_tmp = Image.fromarray(edges).convert("RGB")
-                edges_tmp.save(os.path.join(save_dir, f"edges.png"))
                 edges[edges == 0] = 128
                 return Image.fromarray(edges).convert("RGB")
             c_img = get_canny_edge(c_img)
@@ -210,7 +215,6 @@ def process_image_and_text(condition_image, target_prompt, condition_image_promp
                 )
                 return depth_pipe(img)["depth"].convert("RGB").resize((512, 512))
             c_img = get_depth_map(c_img)
-            c_img.save(os.path.join(save_dir, f"depth.png"))
             k = (255 - 128) / 255
             b = 128
             c_img = c_img.point(lambda x: k * x + b)
@@ -230,7 +234,6 @@ def process_image_and_text(condition_image, target_prompt, condition_image_promp
                 Image.new("RGB", (512, 512), (255, 255, 255)),
                 mask
             )
-            c_img.save(os.path.join(save_dir, f"mask.png"))
             c_img = Image.composite(
                 c_img,
                 Image.new("RGB", (512, 512), (128, 128, 128)),
@@ -238,9 +241,7 @@ def process_image_and_text(condition_image, target_prompt, condition_image_promp
             )
         elif task == "sr":
             c_img = c_img.resize((int(512 / 4), int(512 / 4))).convert("RGB")
-            c_img.save(os.path.join(save_dir, f"low_resolution.png"))
             c_img = c_img.resize((512, 512))
-            c_img.save(os.path.join(save_dir, f"low_to_high.png"))
 
     gen_img = pipe(
         image=c_img,
@@ -318,8 +319,5 @@ def create_app():
 
 
 if __name__ == "__main__":
-    global pipe, current_task
-    pipe = None
-    current_task = None
     init_basemodel()
     create_app().launch(debug=True, ssr_mode=False)