feat: add poster generation from session recap data and improve model caching

app.py

@@ -405,6 +405,38 @@ def generate_recap(session_id: str):
     return "\n".join(lines), result
 
 
+def generate_poster_from_session(session_id: str):
+    """Generate a poster image from an existing session's recap data."""
+    if session_id not in SESSION_STORE:
+        return None, "[!] Unknown session - generate a recap first"
+
+    session = SESSION_STORE[session_id]
+
+    # Check if we have a recap result or can build one
+    if "poster_url" in session:
+        return session["poster_url"], "Poster ready from previous generation"
+
+    # Build story packet from stored data
+    game = session.get("game")
+    if not game:
+        return None, "[!] No game data in session"
+
+    events = load_events(session_id=session_id)
+    journals = load_journal_entries(session_id=session_id)
+    scores = session.get("scores", compute_scores(events, game))
+    photos = session.get("photos", [])
+
+    packet = build_story_packet(
+        game=game, events=events, scores=scores,
+        journal_entries=journals, photo_captions=photos,
+    )
+    result = generate_story(packet, session_id=session_id)
+
+    poster_url, status = generate_poster_sync(session_id, result["poster_prompt"])
+    session["poster_url"] = poster_url
+    return poster_url, status
+
+
 # ── Build summary text ────────────────────────────────────────────────────────
 def build_summary(game: dict, state: dict, session_id: str = "") -> str:
     """Format a human-readable game summary for the UI."""
@@ -617,7 +649,15 @@ with gr.Blocks(title="CityQuest-AI – Game Generator") as demo:
         recap_btn = gr.Button("📖 Generate Recap", variant="primary", size="lg")
         recap_md = gr.Markdown(value="")
         recap_json = gr.JSON(label="Recap data", visible=False)
-
+    #         Poster section
+        gr.Markdown("---\n### Poster")
+        poster_session_id = gr.Textbox(
+            label="Session ID",
+            placeholder="Paste the same Session ID here\u2026",
+        )
+        poster_btn = gr.Button(" Generate Poster", variant="secondary")
+        poster_image = gr.Image(label="Poster", height=400, visible=True)
+        poster_status = gr.Markdown(value="")
     # ── Wire events ────────────────────────────────────────────────────────
 
     # Generate tab
@@ -675,6 +715,12 @@ with gr.Blocks(title="CityQuest-AI – Game Generator") as demo:
         outputs=[recap_md, recap_json],
     )
 
+    poster_btn.click(
+        fn=generate_poster_from_session,
+        inputs=[poster_session_id],
+        outputs=[poster_image, poster_status],
+    )
+
     gr.Markdown(
         """
         ---

app/services/generator.py

@@ -6,8 +6,9 @@ from typing import Optional
 from pathlib import Path
 
 
-# Model initialization cache
+# Model initialization cache (exported for reuse by story.py recap)
 _model_cache = {}
+NEMOTRON_MODEL_CACHE = _model_cache
 
 # Model configuration
 NEMOTRON_MODEL_ID = "nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF"

app/services/story.py

@@ -17,7 +17,7 @@ from pathlib import Path
 from typing import Optional
 
 from app.services.tracing import log_event
-from app.services.generator import NEMOTRON_MODEL_ID, NEMOTRON_GGUF_FILE
+from app.services.generator import NEMOTRON_MODEL_ID, NEMOTRON_GGUF_FILE, NEMOTRON_MODEL_CACHE
 
 
 # ── Story packet builder ───────────────────────────────────────────────────
@@ -248,19 +248,32 @@ def _template_poster_prompt(packet: dict) -> str:
 def _llm_recap(prompt_template: str, story_packet: dict) -> Optional[dict]:
     """Attempt to generate a recap using the Nemotron LLM via llama-cpp-python.
 
+    Reuses the cached model from ``generator._model_cache`` if available so we
+    don't load a second copy of the 2.84 GB model on the same GPU.
+
     Returns None if the model is unavailable so callers can fall back to
     the template engine.
     """
     try:
         from llama_cpp import Llama
 
-        llm = Llama.from_pretrained(
-            repo_id=NEMOTRON_MODEL_ID,
-            filename=NEMOTRON_GGUF_FILE,
-            verbose=False,
-            n_gpu_layers=-1,
-            n_ctx=2048,
-        )
+        # Try to reuse the already-loaded generator model first
+        llm = None
+        for key, cached in NEMOTRON_MODEL_CACHE.items():
+            if key.startswith("llama_cpp_"):
+                llm = cached
+                print("[story] Reusing cached Nemotron model from generator")
+                break
+
+        if llm is None:
+            # Fall back: load a fresh instance
+            llm = Llama.from_pretrained(
+                repo_id=NEMOTRON_MODEL_ID,
+                filename=NEMOTRON_GGUF_FILE,
+                verbose=False,
+                n_gpu_layers=-1,
+                n_ctx=2048,
+            )
 
         packet_str = json.dumps(story_packet, indent=2, default=str)[:3000]
         prompt = prompt_template.format(story_packet=packet_str)
@@ -299,8 +312,9 @@ def _llm_recap(prompt_template: str, story_packet: dict) -> Optional[dict]:
 def generate_story(story_packet: dict, session_id: Optional[str] = None) -> dict:
     """Generate final recap story from game data and events.
 
-    Attempts MiniCPM-based generation first, then Nemotron, then falls
-    back to high-quality templates so the pipeline always returns a result.
+    Attempts Nemotron-based generation first (reuses model already loaded
+    on GPU from game generation), then MiniCPM, then falls back to
+    high-quality templates so the pipeline always returns a result.
 
     Args:
         story_packet: Structured packet with game info, scores, journals, photos.
@@ -310,27 +324,27 @@ def generate_story(story_packet: dict, session_id: Optional[str] = None) -> dict
         Story output dict with keys ``short_recap``, ``long_summary``,
         ``poster_prompt``, ``story_packet``.
     """
-    # Try MiniCPM path first (OpenBMB sponsor model)
+    # Try Nemotron path first (model already loaded on GPU from game generation)
     llm_result = None
-    try:
-        from app.services.minicpm import generate_recap as minicpm_recap
-        llm_result = minicpm_recap(story_packet)
+    template_path = Path("app/prompts/story_recap.txt")
+    prompt_template = ""
+    if template_path.exists():
+        prompt_template = template_path.read_text(encoding="utf-8")
+
+    if prompt_template:
+        llm_result = _llm_recap(prompt_template, story_packet)
         if llm_result:
-            print("[story] MiniCPM recap generated successfully")
-    except Exception as exc:
-        print(f"[story] MiniCPM recap unavailable: {exc}")
+            print("[story] Nemotron recap generated successfully")
 
-    # Fall back to Nemotron LLM path
+    # Fall back to MiniCPM (OpenBMB sponsor model)
     if not llm_result:
-        template_path = Path("app/prompts/story_recap.txt")
-        prompt_template = ""
-        if template_path.exists():
-            prompt_template = template_path.read_text(encoding="utf-8")
-
-        if prompt_template:
-            llm_result = _llm_recap(prompt_template, story_packet)
+        try:
+            from app.services.minicpm import generate_recap as minicpm_recap
+            llm_result = minicpm_recap(story_packet)
             if llm_result:
-                print("[story] Nemotron recap generated successfully")
+                print("[story] MiniCPM recap generated successfully")
+        except Exception as exc:
+            print(f"[story] MiniCPM recap unavailable: {exc}")
 
     if llm_result and all(k in llm_result for k in ("short_recap", "long_summary", "poster_prompt")):
         short_recap = llm_result["short_recap"]