Add generation progress feedback

Surface three phases in the UI via gr.Progress: model loading (first run),
live denoising-step bar (track_tqdm with disable_tqdm off), and safety filter.

app.py

@@ -343,7 +343,10 @@ class T2IEngine:
         micro_cond: str,
         return_animation: bool,
         enable_image_guard: bool,
+        progress: gr.Progress | None = None,
     ):
+        if progress is not None and self._model is None:
+            progress(0.0, desc="Loading model (first run, this can take 1-2 min)…")
         self._lazy_load()
 
         gen_cfg = dict(DEFAULT_GENERATION_CONFIG)
@@ -367,6 +370,9 @@ class T2IEngine:
 
         n_tokens = n_tokens_from_resolution(int(image_resolution))
 
+        if progress is not None:
+            progress(0.0, desc="Generating…")
+
         t0 = time.time()
         with torch.no_grad():
             with torch.inference_mode():
@@ -376,7 +382,7 @@ class T2IEngine:
                     **gen_cfg,
                     image_resolution=int(image_resolution),
                     n_tokens=n_tokens,
-                    disable_tqdm=True,
+                    disable_tqdm=progress is None,
                     return_intermediate_steps=return_animation,
                 )
         latency = time.time() - t0
@@ -391,6 +397,9 @@ class T2IEngine:
             f"edit_threshold={edit_threshold:.3f} | gen_time={latency:.2f}s"
         )
 
+        if progress is not None and enable_image_guard:
+            progress(1.0, desc="Running safety filter…")
+
         guard_ok, guard_meta = self._moderate_output(
             result, return_animation, enable_image_guard
         )
@@ -425,6 +434,7 @@ def generate(
     micro_cond: str,
     return_animation: bool,
     enable_image_guard: bool,
+    progress: gr.Progress = gr.Progress(track_tqdm=True),
 ):
     return engine.generate(
         prompt,
@@ -443,6 +453,7 @@ def generate(
         micro_cond,
         return_animation,
         enable_image_guard,
+        progress=progress,
     )