Use per-mode dynamic ZeroGPU duration (30s fast, 60s high-detail)

Fast mode (1024px) p95 pipeline time is ~14.5s so 30s is safe.
High-Detail mode (2048px) p99 is ~40s so 60s is kept for that path.
Drop mode default params — Gradio always passes it explicitly.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -332,7 +332,7 @@ with open("templates/app.html") as _f:
 
 # ── Gradio blocks ──────────────────────────────────────────────────────────────
 
-def infer(images_b64_json, prompt, seed, randomize_seed, guidance_scale, steps, mode="fast", progress=gr.Progress(track_tqdm=True)):
+def infer(images_b64_json, prompt, seed, randomize_seed, guidance_scale, steps, mode, progress=gr.Progress(track_tqdm=True)):
     # CPU-only preprocessing — GPU not yet allocated
     gc.collect()
     pil_images = b64_to_pil_list(images_b64_json)
@@ -341,7 +341,7 @@ def infer(images_b64_json, prompt, seed, randomize_seed, guidance_scale, steps,
     width, height = update_dimensions_on_upload(pil_images[0], max_dim_for_mode(mode))
     t0 = time.perf_counter()
     try:
-        result_image, seed, duration = _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height)
+        result_image, seed, duration = _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height, mode)
         # _spawn_log is called here (main process) so the thread survives after _infer_gpu's
         # @spaces.GPU subprocess exits — previously the daemon thread was killed on subprocess exit.
         _spawn_log(pil_images, result_image, prompt, seed, steps, guidance_scale, width, height, duration, True)
@@ -352,8 +352,8 @@ def infer(images_b64_json, prompt, seed, randomize_seed, guidance_scale, steps,
         raise
 
 
-@spaces.GPU(duration=60)
-def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height):
+@spaces.GPU(duration=lambda *a, **kw: 30 if (len(a) > 7 and a[7] == "fast") else 60)
+def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height, mode):
     _cuda_ok = torch.cuda.is_available()
     timer = _InferTimer(_cuda_ok)
     t0 = time.perf_counter()