Dedupe reward generation: reuse pre-gen in the on-demand fallback

When the player finished a level before the background pre-generation
completed, the on-demand fallback generated the SAME prompt again — two
generations for one reward (visible in the logs; wasted a ZeroGPU call).

- image_generator: add a single-slot cache keyed by reward_id (level).
generate_reward_image(..., reward_id=level); inside the gen lock, a
matching reward_id returns the cached image instead of regenerating.
A different reward_id always regenerates, so rewards stay varied.
- app.py: pass reward_id=level from both pregenerate_image and
generate_on_demand.

Verified: same reward_id → reused (0.0s, identical image); new id → fresh.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app.py

@@ -198,7 +198,7 @@ def pregenerate_image(state: dict):
         places  = state.get("selected_places",  [random.choice(PLACE_EMOJIS)])
 
         print(f"[pregenerate] level={level} | animals={animals} | places={places}")
-        result, prompt = generate_reward_image(animals, places)
+        result, prompt = generate_reward_image(animals, places, reward_id=level)
         print(f"[pregenerate] level={level} done | prompt={prompt!r}")
 
         if result is not None:
@@ -346,7 +346,7 @@ def generate_on_demand(state: dict, coll_items: list):
         places  = state.get("selected_places",  [random.choice(PLACE_EMOJIS)])
 
         print(f"[on_demand] level={level}")
-        result, prompt = generate_reward_image(animals, places)
+        result, prompt = generate_reward_image(animals, places, reward_id=level)
         print(f"[on_demand] level={level} done")
 
         if result is not None:

image_generator.py

@@ -272,6 +272,10 @@ def build_prompt(animals: list[str], places: list[str]) -> str:
 _pipe = None
 _pipe_lock = threading.Lock()   # serialize loading — pregenerate + on-demand can race
 _gen_lock  = threading.Lock()   # serialize inference — one device can't run two forwards at once
+# Single-slot cache so the on-demand fallback reuses the in-flight pre-generation
+# for the SAME reward instead of generating it twice. Keyed by reward_id (level);
+# a different reward_id always regenerates, so rewards stay varied across levels.
+_recent_reward = {"id": None, "image": None}
 _use_mps = (not IS_HF_SPACE) and (not torch.cuda.is_available()) and torch.backends.mps.is_available()
 _dtype   = torch.float16 if _use_mps else torch.bfloat16   # float16 on MPS (bfloat16 unsupported)
 
@@ -327,7 +331,7 @@ def get_pipeline():
 # Core generation (always wrapped in try/except)
 # ---------------------------------------------------------------------------
 
-def _generate(animals: list[str], places: list[str]):
+def _generate(animals: list[str], places: list[str], reward_id=None):
     try:
         import time
         prompt = build_prompt(animals, places)
@@ -335,6 +339,12 @@ def _generate(animals: list[str], places: list[str]):
         # can fire concurrently, but one device (MPS/GPU slice) can't run two forward
         # passes at once. ZeroGPU serializes @GPU calls for us; locally we must.
         with _gen_lock:
+            # Reuse the pre-generated image for this reward instead of generating twice.
+            if reward_id is not None and _recent_reward["id"] == reward_id \
+                    and _recent_reward["image"] is not None:
+                print(f"♻️  Reusing pre-generated image for reward {reward_id}")
+                return _recent_reward["image"], prompt
+
             pipe = get_pipeline()
             print(f"Generating | prompt: {prompt}")
             t0 = time.time()
@@ -346,7 +356,10 @@ def _generate(animals: list[str], places: list[str]):
                 width=512,
             )
             print(f"✅ Generated in {time.time() - t0:.1f}s")
-        return result.images[0], prompt
+            image = result.images[0]
+            if reward_id is not None:
+                _recent_reward["id"], _recent_reward["image"] = reward_id, image
+        return image, prompt
 
     except Exception as e:
         import traceback
@@ -370,7 +383,8 @@ if IS_HF_SPACE:
 # ---------------------------------------------------------------------------
 
 @GPU(duration=60)
-def generate_reward_image(animals: list[str], places: list[str]):
+def generate_reward_image(animals: list[str], places: list[str], reward_id=None):
     """Generate a reward image.  On HF Spaces runs inside a ZeroGPU slice;
-    locally the @GPU decorator is a no-op and MPS/CPU is used instead."""
-    return _generate(animals, places)
+    locally the @GPU decorator is a no-op and MPS/CPU is used instead.
+    Pass reward_id so the pre-generate + on-demand paths dedupe the same reward."""
+    return _generate(animals, places, reward_id=reward_id)