Update app.py

app.py

@@ -208,8 +208,10 @@ LORA_TRIGGER_WORDS = {
     "ultimate_realistic_breast":"urb, realistic breast",
 }
 
-# Tracks which adapter names have been loaded into the pipeline this session.
-# ZeroGPU resets VRAM on every @spaces.GPU call, so we reload as needed.
+# Tracks which adapter names have been successfully loaded into the pipeline.
+# We keep this across @spaces.GPU calls because the LoRA weights stay in RAM
+# (only active GPU state is reset by ZeroGPU). This avoids re-downloading
+# and re-registering the same adapters on every generation.
 LOADED_ADAPTERS: set[str] = set()
 
 # ── Helpers ────────────────────────────────────────────────────────────────────
@@ -230,10 +232,8 @@ def append_triggers(current_prompt: str, lora_name: str) -> str:
 
 
 def load_and_apply_stack(extra_adapters: list[str], extra_weights: list[float]):
-    """Lazy-load any unseen adapters, then activate the full stack.
-
-    Returns the list of adapters that were successfully loaded (some may be
-    skipped if the remote file is missing or corrupted).
+    """Lazy-load any unseen adapters (only once per app lifetime), then
+    activate exactly the requested stack for this inference.
     """
     if not extra_adapters:
         pipe.disable_lora()
@@ -253,8 +253,14 @@ def load_and_apply_stack(extra_adapters: list[str], extra_weights: list[float]):
                 )
                 LOADED_ADAPTERS.add(name)
             except Exception as e:
-                print(f"WARNING: Failed to load LoRA '{name}': {e}")
-                continue
+                # If it already exists from a previous call in this Space session,
+                # we treat it as success (no need to reload weights).
+                if "already exists" in str(e).lower() or "Adapter" in str(e):
+                    print(f"Adapter '{name}' already registered — reusing.")
+                    LOADED_ADAPTERS.add(name)  # ensure it's tracked
+                else:
+                    print(f"WARNING: Failed to load LoRA '{name}': {e}")
+                    continue
         loaded.append(name)
         weights_out.append(weight)
 
@@ -290,9 +296,9 @@ def infer(
     torch.cuda.empty_cache()
     # ─────────────────────────────────────────────────────────────────────────
 
-    # ── ZeroGPU FIX: GPU state resets between @spaces.GPU calls, so we must
-    #    clear the adapter tracking set to force reloading on each invocation.
-    LOADED_ADAPTERS.clear()
+    # Always start with a clean LoRA state (previous active adapters are
+    # disabled by the finally block, but this is extra safety).
+    pipe.disable_lora()
 
     if input_image is None:
         raise gr.Error("Please upload an image.")
@@ -352,9 +358,8 @@ def infer(
             raise gr.Error(f"GPU error: {e}")
         raise gr.Error(f"Inference failed: {e}")
     finally:
-        # ── OOM FIX: Unload LoRAs and clean up after each inference ──────────
-        if loaded_adapters:
-            pipe.disable_lora()
+        # ── OOM FIX: Unload active LoRAs and clean up after each inference ───
+        pipe.disable_lora()
         gc.collect()
         torch.cuda.empty_cache()
         # ─────────────────────────────────────────────────────────────────────