diff --git a/generate_equirect.py b/generate_equirect.py
index 5f3540d..515e52c 100644
--- a/generate_equirect.py
+++ b/generate_equirect.py
@@ -664,6 +664,14 @@ def generate(
                     "install accelerate for low_cpu_mem_usage: pip install accelerate. "
                     f"Errors: {load_errors}"
                 ) from (last_err or Exception("No pipeline loaded"))
+        if gen_pipe is None:
+            mode = "offline/local-cache" if local_files_only() else "online"
+            raise RuntimeError(
+                f"Failed to load model '{model_path}' in {mode} mode. "
+                "The local single-file checkpoint needs the bundled SDXL config and locally cached "
+                "tokenizer/text-encoder metadata. "
+                f"Errors: {load_errors}"
+            )
         if gen_pipe.vae is None and is_sdxl:
             try:
                 vae = vae or AutoencoderKL.from_pretrained(