Update app.py

app.py

@@ -1,44 +1,40 @@
 import os
-
-if os.getcwd() != '/home/user/app':
-    os.chdir('/home/user/app')
-
 import sys
-import spaces
 import subprocess
 import asyncio
+import uuid
 from typing import Sequence, Mapping, Any, Union
 
+# --- 2. Let ComfyUI's main.py handle all initial setup ---
 print("Importing ComfyUI's main.py for setup...")
 import main
 print("ComfyUI main imported.")
 
-
+# --- 3. Now we can import the rest of the necessary modules ---
 import torch
 import gradio as gr
 from huggingface_hub import hf_hub_download
 from comfy import model_management
+import spaces
 from PIL import Image
 import random
-import nodes # Import nodes after main has set everything up
-
+import nodes
 
-# --- Manually trigger the node initialization ---
-# This step is normally done inside main.start_comfyui(), but we do it here.
-# It loads all built-in, extra, and custom nodes into the NODE_CLASS_MAPPINGS.
+# --- 4. Manually trigger the node initialization ---
 print("Initializing ComfyUI nodes...")
 loop = asyncio.new_event_loop()
 asyncio.set_event_loop(loop)
 loop.run_until_complete(nodes.init_extra_nodes())
 print("Nodes initialized.")
 
-# --- Helper function from the original script ---
+# --- Helper function ---
 def get_value_at_index(obj: Union[Sequence, Mapping], index: int) -> Any:
     try:
         return obj[index]
     except KeyError:
         return obj["result"][index]
 
+
 # --- Model Downloads ---
 print("Downloading models from Hugging Face Hub...")
 hf_hub_download(repo_id="Comfy-Org/Wan_2.1_ComfyUI_repackaged", filename="split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors", local_dir="models/text_encoders")
@@ -52,7 +48,6 @@ print("Downloads complete.")
 
 
 # --- ZeroGPU: Pre-load models and instantiate nodes globally ---
-# This part will now work because NODE_CLASS_MAPPINGS is correctly populated.
 cliploader = nodes.NODE_CLASS_MAPPINGS["CLIPLoader"]()
 cliptextencode = nodes.NODE_CLASS_MAPPINGS["CLIPTextEncode"]()
 unetloader = nodes.NODE_CLASS_MAPPINGS["UNETLoader"]()
@@ -68,7 +63,6 @@ ksampleradvanced = nodes.NODE_CLASS_MAPPINGS["KSamplerAdvanced"]()
 vaedecode = nodes.NODE_CLASS_MAPPINGS["VAEDecode"]()
 createvideo = nodes.NODE_CLASS_MAPPINGS["CreateVideo"]()
 savevideo = nodes.NODE_CLASS_MAPPINGS["SaveVideo"]()
-imageresize = nodes.NODE_CLASS_MAPPINGS["ImageResize+"]()
 
 cliploader_38 = cliploader.load_clip(clip_name="umt5_xxl_fp8_e4m3fn_scaled.safetensors", type="wan", device="cpu")
 unetloader_37_low_noise = unetloader.load_unet(unet_name="wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors", weight_dtype="default")
@@ -88,8 +82,7 @@ valid_models = [getattr(loader[0], 'patcher', loader[0]) for loader in model_loa
 model_management.load_models_gpu(valid_models)
 
 # --- App Logic ---
-def calculate_dimensions(image_path):
-    with Image.open(image_path) as img: width, height = img.size
+def calculate_dimensions(width, height):
     if width == height: return 480, 480
     if width > height: new_width, new_height = 832, int(height * (832 / width))
     else: new_height, new_width = 832, int(width * (832 / height))
@@ -97,23 +90,46 @@ def calculate_dimensions(image_path):
 
 @spaces.GPU(duration=120)
 def generate_video(prompt, first_image_path, last_image_path, duration_seconds):
+    # Create a temporary directory for resized images
+    temp_dir = f"temp_resized_{uuid.uuid4().hex}"
+    os.makedirs(temp_dir, exist_ok=True)
+    
     with torch.inference_mode():
+        # --- Python Image Preprocessing using Pillow ---
+        print("Preprocessing images with Pillow...")
+        with Image.open(first_image_path) as img:
+            orig_width, orig_height = img.size
+        
+        target_width, target_height = calculate_dimensions(orig_width, orig_height)
+        
+        # Resize first image
+        with Image.open(first_image_path) as img:
+            img_resized = img.resize((target_width, target_height), Image.Resampling.LANCZOS)
+            resized_first_path = os.path.join(temp_dir, "first_frame_resized.png")
+            img_resized.save(resized_first_path)
+
+        # Resize second image to match the target dimensions
+        with Image.open(last_image_path) as img:
+            img_resized = img.resize((target_width, target_height), Image.Resampling.LANCZOS)
+            resized_last_path = os.path.join(temp_dir, "last_frame_resized.png")
+            img_resized.save(resized_last_path)
+        print(f"Images resized to {target_width}x{target_height} and saved temporarily.")
+        # --- End Preprocessing ---
+
         FPS, MAX_FRAMES = 16, 81
         length_in_frames = max(1, min(int(duration_seconds * FPS), MAX_FRAMES))
         print(f"Requested duration: {duration_seconds}s. Calculated frames: {length_in_frames}")
-        target_width, target_height = calculate_dimensions(first_image_path)
-
-        loaded_first_image = loadimage.load_image(image=first_image_path)
-        resized_first_image = imageresize.execute(width=target_width, height=target_height, interpolation="bicubic", method="stretch", image=get_value_at_index(loaded_first_image, 0))
-        loaded_last_image = loadimage.load_image(image=last_image_path)
-        resized_last_image = imageresize.execute(width=target_width, height=target_height, interpolation="bicubic", method="stretch", image=get_value_at_index(loaded_last_image, 0))
-
+        
+        # Load the pre-processed images into ComfyUI
+        loaded_first_image = loadimage.load_image(image=os.path.basename(resized_first_path))
+        loaded_last_image = loadimage.load_image(image=os.path.basename(resized_last_path))
+        
         cliptextencode_6 = cliptextencode.encode(text=prompt, clip=get_value_at_index(cliploader_38, 0))
         cliptextencode_7_negative = cliptextencode.encode(text="low quality, worst quality, jpeg artifacts, ugly, deformed, blurry", clip=get_value_at_index(cliploader_38, 0))
-        clipvisionencode_51 = clipvisionencode.encode(crop="none", clip_vision=get_value_at_index(clipvisionloader_49, 0), image=get_value_at_index(resized_first_image, 0))
-        clipvisionencode_87 = clipvisionencode.encode(crop="none", clip_vision=get_value_at_index(clipvisionloader_49, 0), image=get_value_at_index(resized_last_image, 0))
+        clipvisionencode_51 = clipvisionencode.encode(crop="none", clip_vision=get_value_at_index(clipvisionloader_49, 0), image=get_value_at_index(loaded_first_image, 0))
+        clipvisionencode_87 = clipvisionencode.encode(crop="none", clip_vision=get_value_at_index(clipvisionloader_49, 0), image=get_value_at_index(loaded_last_image, 0))
 
-        wanfirstlastframetovideo_83 = wanfirstlastframetovideo.EXECUTE_NORMALIZED(width=target_width, height=target_height, length=length_in_frames, batch_size=1, positive=get_value_at_index(cliptextencode_6, 0), negative=get_value_at_index(cliptextencode_7_negative, 0), vae=get_value_at_index(vaeloader_39, 0), clip_vision_start_image=get_value_at_index(clipvisionencode_51, 0), clip_vision_end_image=get_value_at_index(clipvisionencode_87, 0), start_image=get_value_at_index(resized_first_image, 0), end_image=get_value_at_index(resized_last_image, 0))
+        wanfirstlastframetovideo_83 = wanfirstlastframetovideo.EXECUTE_NORMALIZED(width=target_width, height=target_height, length=length_in_frames, batch_size=1, positive=get_value_at_index(cliptextencode_6, 0), negative=get_value_at_index(cliptextencode_7_negative, 0), vae=get_value_at_index(vaeloader_39, 0), clip_vision_start_image=get_value_at_index(clipvisionencode_51, 0), clip_vision_end_image=get_value_at_index(clipvisionencode_87, 0), start_image=get_value_at_index(loaded_first_image, 0), end_image=get_value_at_index(loaded_last_image, 0))
 
         ksampler_positive = get_value_at_index(wanfirstlastframetovideo_83, 0)
         ksampler_negative = get_value_at_index(wanfirstlastframetovideo_83, 1)
@@ -128,7 +144,7 @@ def generate_video(prompt, first_image_path, last_image_path, duration_seconds):
         
         return f"output/{savevideo_103['ui']['videos'][0]['filename']}"
 
-# --- Gradio Interface (no changes needed) ---
+# --- Gradio Interface ---
 with gr.Blocks() as app:
     gr.Markdown("# Wan 2.2 First/Last Frame to Video")
     gr.Markdown("Provide a starting image, an ending image, a text prompt, and a desired duration to generate a video transitioning between them.")
@@ -149,4 +165,7 @@ if __name__ == "__main__":
     if not os.path.exists("examples"): os.makedirs("examples")
     if not os.path.exists("examples/start.png"): Image.new('RGB', (512, 512), color='red').save('examples/start.png')
     if not os.path.exists("examples/end.png"): Image.new('RGB', (512, 512), color='blue').save('examples/end.png')
+    # Set the input directory for LoadImage to find the temp files
+    import folder_paths
+    folder_paths.add_model_folder_path("input", "temp_resized")
     app.launch()