y

requirements.txt

@@ -1,12 +1,7 @@
-# Gradio UI
 gradio
-# Transformers + Diffusers
 transformers>=4.30.0
 diffusers>=0.16.0
-# PyTorch
 torch>=2.4.0
-# Wan-VAE and DiT dependencies
 xfuser>=0.4.1
 accelerate
-# Optional: huggingface_hub for model download
 huggingface_hub>=0.13.0

app.py

@@ -1,32 +1,38 @@
+import os
 import gradio as gr
 import torch
 from diffusers import DiffusionPipeline
-from huggingface_hub import hf_hub_download
 
-# Load model weights from Hub
+# Read token from environment (configured as a Space secret)
+token = os.environ.get("HUGGINGFACE_TOKEN")
+if token is None:
+    raise ValueError("Environment variable HUGGINGFACE_TOKEN is not set.")
+
 model_id = "Wan-AI/Wan2.1-I2V-14B-480P"
-ckpt_dir = hf_hub_download(repo_id=model_id, filename=".")
 
+# Load pipeline directly from the Hub, using the token
 pipe = DiffusionPipeline.from_pretrained(
     model_id,
     torch_dtype=torch.float16,
-    use_auth_token=True
+    trust_remote_code=True,
+    use_auth_token=token
 ).to("cuda")
 
+# Enable memory-saving features
 pipe.enable_attention_slicing()
 
-
-def generate_video(image, prompt, num_frames=16):
-    video = pipe(
+# Generation function
+def generate_video(image, prompt, num_frames=16, steps=50, guidance_scale=7.5):
+    result = pipe(
         prompt=prompt,
         init_image=image,
-        num_inference_steps=50,
-        guidance_scale=7.5,
+        num_inference_steps=steps,
+        guidance_scale=guidance_scale,
         num_frames=num_frames
-    ).videos
-    return video
+    )
+    return result.videos
 
-# Gradio UI
+# Gradio UI definition
 def main():
     with gr.Blocks() as demo:
         gr.Markdown("# Wan2.1 Image-to-Video Demo")