Update app.py

app.py

@@ -16,8 +16,7 @@ import cv2
 from transformers import (
     Qwen2_5_VLForConditionalGeneration,
     Qwen2VLForConditionalGeneration,
-    Gemma3ForConditionalGeneration,
-    LlavaForConditionalGeneration,
+    Llama4ForConditionalGeneration,
     AutoProcessor,
     TextIteratorStreamer,
 )
@@ -30,10 +29,10 @@ MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
-# Load gemma-3-4b-it
-MODEL_ID_M = "google/gemma-3-4b-it"
+# Load meta-llama/Llama-Guard-4-12B
+MODEL_ID_M = "meta-llama/Llama-Guard-4-12B"
 processor_m = AutoProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
-model_m = Gemma3ForConditionalGeneration.from_pretrained(
+model_m = Llama4ForConditionalGeneration.from_pretrained(
     MODEL_ID_M,
     trust_remote_code=True,
     torch_dtype=torch.float16
@@ -49,14 +48,25 @@ model_x = Qwen2VLForConditionalGeneration.from_pretrained(
 ).to(device).eval()
 
 # Load Relaxed
-MODEL_ID_Z = "fancyfeast/llama-joycaption-beta-one-hf-llava"
+MODEL_ID_Z = "Qwen/Qwen2.5-VL-3B-Instruct"
 processor_z = AutoProcessor.from_pretrained(MODEL_ID_Z, trust_remote_code=True)
-model_z = LlavaForConditionalGeneration.from_pretrained(
+model_z = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     MODEL_ID_Z,
     trust_remote_code=True,
     torch_dtype=torch.float16
 ).to(device).eval()
 
+# Load ImageScope
+MODEL_ID_T = "prithivMLmods/Imgscope-OCR-2B-0527"
+processor_t = AutoProcessor.from_pretrained(MODEL_ID_X, trust_remote_code=True)
+model_t = Qwen2VLForConditionalGeneration.from_pretrained(
+    MODEL_ID_T,
+    trust_remote_code=True,
+    torch_dtype=torch.float16
+).to(device).eval()
+
+
+
 def downsample_video(video_path):
     """
     Downsamples the video to evenly spaced frames.
@@ -88,15 +98,18 @@ def generate_image(model_name: str, text: str, image: Image.Image,
     """
     Generates responses using the selected model for image input.
     """
-    if model_name == "gemma-3-4b-it":
+    if model_name == "Llama-4":
         processor = processor_m
         model = model_m
     elif model_name == "coreOCR-7B-050325-preview":
         processor = processor_x
         model = model_x
-    elif model_name == "llama-joycaption":
+    elif model_name == "Qwen2.5-VL-3B":
         processor = processor_z
         model = model_z
+    elif model_name == "Imgscope-OCR-2B":
+        processor = processor_t
+        model = model_t
     else:
         yield "Invalid model selected."
         return
@@ -141,15 +154,18 @@ def generate_video(model_name: str, text: str, video_path: str,
     """
     Generates responses using the selected model for video input.
     """
-    if model_name == "gemma-3-4b-it":
+    if model_name == "Llama-4":
         processor = processor_m
         model = model_m
     elif model_name == "coreOCR-7B-050325-preview":
         processor = processor_x
         model = model_x
-    elif model_name == "llama-joycaption":
+    elif model_name == "Qwen2.5-VL-3B":
         processor = processor_z
         model = model_z
+    elif model_name == "Imgscope-OCR-2B":
+        processor = processor_t
+        model = model_t
     else:
         yield "Invalid model selected."
         return
@@ -247,9 +263,9 @@ with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
         with gr.Column():
             output = gr.Textbox(label="Output", interactive=False, lines=2, scale=2)
             model_choice = gr.Radio(
-                choices=["gemma-3-4b-it", "coreOCR-7B-050325-preview", "llama-joycaption"],
+                choices=["Llama-4", "coreOCR-7B-050325-preview", "Imgscope-OCR-2B", "Qwen2.5-VL-3B"],
                 label="Select Model",
-                value="gemma-3-4b-it"
+                value="Llama-4"
             )
             
     image_submit.click(
@@ -264,4 +280,4 @@ with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
     )
 
 if __name__ == "__main__":
-    demo.queue(max_size=30).launch(share=True, mcp_server=True, ssr_mode=False, show_error=True)
+    demo.queue(max_size=40).launch(share=True, mcp_server=True, ssr_mode=False, show_error=True)