different examples

app.py

@@ -8,7 +8,7 @@ login(token=os.environ["HUGGINGFACE_TOKEN"])
 
 demo_imgs = [
     ["images/chinchilla_web-1024x683.jpg", "images/shiba-inu-dog-in-the-snow.jpg", "images/900.jpeg", "images/dogs.jpeg"],
-    ["images/900.jpeg", "images/hummus.jpg", "images/london-underground-sign.jpg", "images/COCO_train2014_000000194806.jpg"],
+    ["images/hummus.jpg", "images/london-underground-sign.jpg", "images/4645808729_2dfc59b6a5_z.jpg", "images/5944609705_4664531909_z.jpg"],
     ["images/COCO_train2014_000000572279.jpg", "images/COCO_train2014_000000194806.jpg", "images/istockphoto-622434332-1024x1024.jpg", "images/11887_pesto-pasta_Rita-1x1-1-501c953b29074ab193e2b5ad36e64648.jpg"],
     [
         "images/bcee7a-20190225-a-london-underground-sign.jpg",
@@ -25,10 +25,10 @@ demo_texts = [
         "Output: These are labrador retrievers. They are found in the UK.",
     ],
     [
-        "Output: a pink flamingo standing in a body of water.",
         "Output: a bowl filled with creamy hummus placed on a white countertop.",
-        "Output: a red and blue 'Underground' sign found in London",
-        "Output: Two people flying a white kite on a beach."
+        "Output: a red and blue 'Underground' sign found in London.",
+        "Output: a man and a woman on a train looking at their cell phones.",
+        "Output: a lavish reception room with black and white tiled floor."
     ],
     [
         "Question: Describe the scene. Answer: A white airplane being repaired on the runway. 'Cargo' is written on it in red.",
@@ -156,7 +156,7 @@ def generate(
         vision_x=vision_x,
         lang_x=input_ids,
         attention_mask=attention_mask,
-        max_new_tokens=50,
+        max_new_tokens=30,
         num_beams=5,
         do_sample=True,
         top_p=0.95,
@@ -199,6 +199,7 @@ with gr.Blocks() as demo:
     # 🦩 OpenFlamingo Demo  
     
     Blog posts: #1 [An open-source framework for training vision-language models with in-context learning](https://laion.ai/blog/open-flamingo/) // #2 [OpenFlamingo v2: New Models and Enhanced Training Setup]()
+    
     GitHub: [open_flamingo](https://github.com/mlfoundations/open_flamingo)
     
     In this demo we implement an interactive interface that showcases the in-context learning capabilities of the OpenFlamingo-4B model, a large multimodal model trained on top of RedPajama-INCITE-Instruct-3B-v1.