Upload app.py with huggingface_hub

app.py

@@ -2,7 +2,6 @@ import random
 import torch
 import gradio as gr
 import spaces
-from diffusers import FlowMatchEulerDiscreteScheduler, QwenImageEditPlusPipeline
 from PIL import Image
 
 # ── Constantes poses ──────────────────────────────────────────────────────────
@@ -21,29 +20,6 @@ ELEVATION_NAMES = {
 }
 DISTANCE_NAMES = {0.6: "close-up", 1.0: "medium shot", 1.8: "wide shot"}
 
-# ── Chargement modèle ─────────────────────────────────────────────────────────
-
-dtype = torch.bfloat16
-device = "cuda" if torch.cuda.is_available() else "cpu"
-
-pipe = QwenImageEditPlusPipeline.from_pretrained(
-    "Qwen/Qwen-Image-Edit-2511",
-    torch_dtype=dtype,
-)
-pipe.to(device)
-
-pipe.load_lora_weights(
-    "lightx2v/Qwen-Image-Edit-2511-Lightning",
-    weight_name="Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors",
-    adapter_name="lightning",
-)
-pipe.load_lora_weights(
-    "fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA",
-    weight_name="qwen-image-edit-2511-multiple-angles-lora.safetensors",
-    adapter_name="angles",
-)
-pipe.set_adapters(["lightning", "angles"], adapter_weights=[1.0, 1.0])
-
 # ── Helpers ───────────────────────────────────────────────────────────────────
 
 def snap_to_nearest(value, options):
@@ -56,9 +32,39 @@ def build_camera_prompt(azimuth: float, elevation: float, distance: float) -> st
     di = snap_to_nearest(distance, DISTANCES)
     return f"<sks> {AZIMUTH_NAMES[az]} {ELEVATION_NAMES[el]} {DISTANCE_NAMES[di]}"
 
+
+def update_prompt_preview(azimuth, elevation, distance):
+    return build_camera_prompt(azimuth, elevation, distance)
+
+# ── Chargement modèle (lazy, sur GPU) ────────────────────────────────────────
+
+pipe = None
+
+def get_pipe():
+    global pipe
+    if pipe is not None:
+        return pipe
+    from diffusers import QwenImageEditPlusPipeline
+    pipe = QwenImageEditPlusPipeline.from_pretrained(
+        "Qwen/Qwen-Image-Edit-2511",
+        torch_dtype=torch.bfloat16,
+    ).to("cuda")
+    pipe.load_lora_weights(
+        "lightx2v/Qwen-Image-Edit-2511-Lightning",
+        weight_name="Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors",
+        adapter_name="lightning",
+    )
+    pipe.load_lora_weights(
+        "fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA",
+        weight_name="qwen-image-edit-2511-multiple-angles-lora.safetensors",
+        adapter_name="angles",
+    )
+    pipe.set_adapters(["lightning", "angles"], adapter_weights=[1.0, 1.0])
+    return pipe
+
 # ── Inférence ZeroGPU ─────────────────────────────────────────────────────────
 
-@spaces.GPU
+@spaces.GPU(duration=120)
 def infer(
     image: Image.Image,
     azimuth: float,
@@ -70,10 +76,11 @@ def infer(
     if randomize_seed:
         seed = random.randint(0, 2**31 - 1)
 
+    p = get_pipe()
     prompt = build_camera_prompt(azimuth, elevation, distance)
-    generator = torch.Generator(device=device).manual_seed(seed)
+    generator = torch.Generator(device="cuda").manual_seed(seed)
 
-    result = pipe(
+    result = p(
         image=[image],
         prompt=prompt,
         height=1024,
@@ -88,10 +95,6 @@ def infer(
 
 # ── UI Gradio ─────────────────────────────────────────────────────────────────
 
-def update_prompt_preview(azimuth, elevation, distance):
-    return build_camera_prompt(azimuth, elevation, distance)
-
-
 with gr.Blocks(title="Angle Studio", theme=gr.themes.Base()) as demo:
     gr.Markdown(
         """
@@ -142,7 +145,6 @@ with gr.Blocks(title="Angle Studio", theme=gr.themes.Base()) as demo:
 
     session_images = gr.State([])
 
-    # Mise à jour du prompt en temps réel
     for slider in [azimuth_slider, elevation_slider, distance_slider]:
         slider.change(
             fn=update_prompt_preview,