Update app.py

app.py

@@ -24,10 +24,12 @@ pipe.scheduler = DPMSolverMultistepScheduler.from_config(
     use_karras_sigmas=True,
 )
 pipe.enable_attention_slicing()
+
+# Keep on CPU at startup — ZeroGPU allocates GPU per request
 print("Pipeline ready.")
 
 # ── Generation ────────────────────────────────────────────────────────────────
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=180)
 def generate(prompt, negative_prompt, width, height, steps, guidance, seed, randomize):
     if not prompt.strip():
         raise gr.Error("Please enter a prompt.")
@@ -35,7 +37,9 @@ def generate(prompt, negative_prompt, width, height, steps, guidance, seed, rand
         seed = random.randint(0, 2**32 - 1)
     seed = int(seed)
 
-    # CPU generator works correctly with ZeroGPU — do not hardcode "cuda"
+    # Move to GPU for this request
+    pipe.to("cuda")
+
     generator = torch.Generator(device="cpu").manual_seed(seed)
 
     result = pipe(
@@ -48,6 +52,10 @@ def generate(prompt, negative_prompt, width, height, steps, guidance, seed, rand
         generator=generator,
         clip_skip=2,
     )
+
+    # Move back to CPU to free GPU for other users
+    pipe.to("cpu")
+
     return result.images[0], seed
 
 # ── CSS ───────────────────────────────────────────────────────────────────────
@@ -206,11 +214,6 @@ input[type=checkbox] {
     width: 18px !important;
     height: 18px !important;
 }
-.gradio-checkbox label span {
-    color: #9966cc !important;
-    font-size: 0.82em !important;
-    font-weight: 600 !important;
-}
 
 label span {
     color: #8855bb !important;