Update app.py

app.py

@@ -1,5 +1,3 @@
-# third
-
 """
 3D Camera View Generator
 - Qwen Image Edit + Lightning LoRA + Multi-Angle LoRA
@@ -34,41 +32,36 @@ MAX_SEED = np.iinfo(np.int32).max
 dtype   = torch.bfloat16
 device  = "cuda" if torch.cuda.is_available() else "cpu"
 
-# ── Model (lazy-loaded inside @spaces.GPU) ─────────────────────────────────────
+# ── Model Loading ──────────────────────────────────────────────────────────────
 #
 # HOW MODEL LOADING WORKS ON ZEROGPU
 # ────────────────────────────────────
-# • Weights are downloaded to the HF cache on disk the first time
-#   from_pretrained() is called (happens inside the @spaces.GPU window).
-# • They are NOT loaded into GPU VRAM at app startup — only when the first
-#   @spaces.GPU-decorated function actually runs.
-# • After the first call `pipe` stays alive in CPU process memory, so
-#   subsequent requests skip the download and just run inference.
-# • ZeroGPU allocates an H200 for the duration of the decorated function
-#   then releases it, so nothing holds the GPU between user requests.
+# • Model is loaded at module-level (app startup), NOT lazily on first request.
+#   This means weights are deserialised from disk and resident in CPU memory
+#   before any user ever hits the space.
+# • When @spaces.GPU fires, ZeroGPU pins those already-resident CPU tensors
+#   to CUDA — a fast host→device transfer, not a disk read.
+# • Lazy loading (inside @spaces.GPU) would incur disk I/O + deserialisation
+#   on the first call, adding several extra seconds to the cold start.
+# • Between calls the GPU is released by ZeroGPU, but the weights stay in
+#   CPU memory so the next call only pays the H2D transfer cost.
 #
-pipe = None
-
-def load_model():
-    global pipe
-    if pipe is None:
-        pipe = QwenImageEditPlusPipeline.from_pretrained(
-            "Qwen/Qwen-Image-Edit-2511",
-            torch_dtype=dtype,
-        ).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])
-    return pipe
+pipe = QwenImageEditPlusPipeline.from_pretrained(
+    "Qwen/Qwen-Image-Edit-2511",
+    torch_dtype=dtype,
+).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])
 
 
 # ── Camera parameter tables ────────────────────────────────────────────────────
@@ -134,9 +127,7 @@ def infer_camera_edit(
 
     generator = torch.Generator(device=device).manual_seed(seed)
     prompt    = build_camera_prompt(azimuth, elevation, distance)
-    model     = load_model()
-
-    result = model(
+    result = pipe(
         image=image,
         prompt=prompt,
         height=height,
@@ -513,7 +504,7 @@ def create_app():
                 with gr.Accordion("⚙  Generation Settings", open=False):
                     seed_slider  = gr.Slider(0, MAX_SEED, value=42,  step=1,   label="Seed")
                     rand_seed_cb = gr.Checkbox(True, label="Randomise seed each generation")
-                    guidance_sl  = gr.Slider(1.0, 20.0, value=7.5,  step=0.1, label="Guidance Scale")
+                    guidance_sl  = gr.Slider(1.0, 20.0, value=1.0,  step=0.1, label="Guidance Scale (keep ≤1 for Lightning LoRA)")
                     steps_sl     = gr.Slider(1,   50,   value=4,    step=1,   label="Inference Steps")
                     width_sl     = gr.Slider(256, 1024, value=1024, step=32,  label="Width (px)")
                     height_sl    = gr.Slider(256, 1024, value=1024, step=32,  label="Height (px)")