Update app.py

app.py

@@ -12,6 +12,7 @@ Run:
     python app.py
 """
 
+import spaces  # MUST be imported before torch/anything CUDA-related, ZeroGPU requirement
 import os
 import torch
 import gradio as gr
@@ -19,27 +20,21 @@ from diffusers import StableDiffusionXLInpaintPipeline
 from PIL import Image
 
 MODEL_ID = "ShinoharaHare/Waifu-Inpaint-XL"
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32
+DTYPE = torch.float16
 
-pipe = None  # lazy-loaded so the UI opens instantly, model loads on first click
-
-
-def load_pipeline():
-    global pipe
-    if pipe is None:
-        pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
-            MODEL_ID,
-            torch_dtype=DTYPE,
-            use_safetensors=True,
-        )
-        pipe.to(DEVICE)
-        if DEVICE == "cuda":
-            pipe.enable_vae_slicing()
-            pipe.enable_attention_slicing()
-    return pipe
+# Load once at startup. Moving to 'cuda' here is fine under ZeroGPU -- the actual
+# GPU device is only allocated when a @spaces.GPU-decorated function is called.
+pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
+    MODEL_ID,
+    torch_dtype=DTYPE,
+    use_safetensors=True,
+)
+pipe.to("cuda")
+pipe.enable_vae_slicing()
+pipe.enable_attention_slicing()
 
 
+@spaces.GPU(duration=60)  # seconds of GPU time requested per call; raise if you increase steps/variations
 def run_inpaint(
     editor_value,          # gr.ImageEditor output: {"background":..., "layers":[...], "composite":...}
     prompt,
@@ -60,13 +55,11 @@ def run_inpaint(
     mask_layer = editor_value["layers"][0]
     mask = mask_layer.split()[-1].convert("L")  # alpha channel -> grayscale mask
 
-    p = load_pipeline()
-
     results = []
     base_seed = int(seed) if seed >= 0 else torch.seed()
     for i in range(int(num_variations)):
-        gen = torch.Generator(device=DEVICE).manual_seed(base_seed + i)
-        out = p(
+        gen = torch.Generator(device="cuda").manual_seed(base_seed + i)
+        out = pipe(
             prompt=prompt,
             negative_prompt=negative_prompt or None,
             image=base_image,
@@ -112,5 +105,5 @@ with gr.Blocks(title="Waifu-Inpaint-XL") as demo:
     )
 
 if __name__ == "__main__":
-    # share=True gives you a public URL for free when running on Colab/Kaggle
-    demo.launch(share=os.environ.get("GRADIO_SHARE", "true").lower() == "true")
+    # Spaces already serves a public URL -- do NOT pass share=True here (errors on Spaces).
+    demo.launch()