fix: keep vae

app.py

@@ -239,7 +239,12 @@ def generate(
 # Model initialization
 if torch.cuda.is_available():
     try:
-        pipe = utils.load_pipeline(MODEL, device)
+        logger.info("Loading VAE and pipeline...")
+        vae = AutoencoderKL.from_pretrained(
+            "madebyollin/sdxl-vae-fp16-fix",
+            torch_dtype=torch.float16,
+        )
+        pipe = utils.load_pipeline(MODEL, device, vae=vae)
         logger.info("Pipeline loaded successfully on GPU!")
     except Exception as e:
         logger.error(f"Error loading VAE, falling back to default: {e}")