Per-resolution GPU size: 512 → large (1× quota), 1024 → xlarge (2× quota); plain dispatcher routes between two @spaces.GPU entrypoints

app.py

@@ -236,8 +236,7 @@ import threading
 import queue as _queue
 
 
-@spaces.GPU(duration=90, size="xlarge")
-def generate(
+def _generate_core(
     prompt: str,
     num_inference_steps: int = 28,
     guidance_scale: float = 5.0,
@@ -340,6 +339,23 @@ def generate(
     )
 
 
+# Two decorated entrypoints: 1024 needs the full 96GB Blackwell, 512 fits the MIG.
+# `size` isn't dynamic via the decorator, so we route per request from a plain dispatcher.
+@spaces.GPU(duration=60)
+def generate_large(*args, **kwargs):
+    yield from _generate_core(*args, **kwargs)
+
+
+@spaces.GPU(duration=90, size="xlarge")
+def generate_xlarge(*args, **kwargs):
+    yield from _generate_core(*args, **kwargs)
+
+
+def generate(prompt, num_inference_steps, guidance_scale, seed, resolution, randomize_seed):
+    fn = generate_xlarge if int(resolution) >= 1024 else generate_large
+    yield from fn(prompt, num_inference_steps, guidance_scale, seed, resolution, randomize_seed)
+
+
 DESCRIPTION = """
 # 🪄 PiD — Pixel Diffusion Decoder for Z-Image