Update app.py

app.py

@@ -1,4 +1,4 @@
-# app.py – minimal, CPU‑only, high‑quality captions
+# app.py – corrected for Gradio 6+
 import gradio as gr
 from PIL import Image
 import requests, urllib.parse, threading, time
@@ -30,7 +30,7 @@ rewriter = T5ForConditionalGeneration.from_pretrained(TXT_MODEL).to(device).eval
 # Helpers
 # -------------------------------------------------
 def load_image(url: str):
-    """Return PIL image or (None, error). Handles http/https and data‑URL."""
+    """Return (PIL.Image, None) or (None, error). Handles http/https and data‑URL."""
     try:
         url = url.strip()
         if url.startswith("data:"):
@@ -63,7 +63,6 @@ def generate_base(img: Image.Image, max_len=40, beams=2, sample=False):
             early_stopping=True,
         )
     else:
-        # ensure num_return ≤ beams
         out = vision.generate(
             pix,
             max_length=max_len,
@@ -72,8 +71,7 @@ def generate_base(img: Image.Image, max_len=40, beams=2, sample=False):
             early_stopping=True,
         )
     caps = [tokenizer.decode(o, skip_special_tokens=True).strip() for o in out]
-    # pick longest (most detailed)
-    return max(caps, key=lambda s: len(s.split()))
+    return max(caps, key=lambda s: len(s.split()))  # longest = most detailed
 
 def expand_caption(base: str, prompt: str = None, max_len=160):
     """Rich T5 expansion."""
@@ -151,15 +149,21 @@ def final_caption(url, prompt, detail, beams, sample):
 # UI
 # -------------------------------------------------
 css = "footer {display:none !important;}"
-with gr.Blocks(css=css, title="Image Describer (CPU)") as demo:
-    gr.Markdown("## Image Describer – fast base caption + optional detailed rewrite")
+with gr.Blocks() as demo:                     # no css here
+    gr.Markdown(
+        "## Image Describer"
+    )
     with gr.Row():
         with gr.Column():
             url_in = gr.Textbox(label="Image URL / data‑URL")
             prompt_in = gr.Textbox(label="Optional prompt")
-            detail_in = gr.Radio(["Low", "Medium", "High"], value="Medium", label="Detail level")
-            beams_in = gr.Slider(1, 4, step=1, value=2, label="Beams (higher = better, slower)")
-            sample_in = gr.Checkbox(label="Enable sampling (more diverse)", value=False)
+            detail_in = gr.Radio(
+                ["Low", "Medium", "High"], value="Medium", label="Detail level"
+            )
+            beams_in = gr.Slider(1, 4, step=1, value=2, label="Beams")
+            sample_in = gr.Checkbox(
+                label="Enable sampling (more diverse)", value=False
+            )
             go_btn = gr.Button("Load & Describe (fast)")
             final_btn = gr.Button("Get final caption (detailed)")
             status_out = gr.Textbox(label="Status", interactive=False)
@@ -179,5 +183,15 @@ with gr.Blocks(css=css, title="Image Describer (CPU)") as demo:
         outputs=[caption_out, status_out],
     )
 
+# -------------------------------------------------
+# Launch – css and title are passed here (Gradio 6+)
+# -------------------------------------------------
 if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860)
+    demo.queue()  # enables background threads without leaking the event loop
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        css=css,
+        title="Image Describer (CPU)",
+        prevent_thread_lock=True,   # clean shutdown → no “invalid file descriptor” warnings
+    )