Upload 2 files

app.py

@@ -1,68 +1,156 @@
 import gradio as gr
 from PIL import Image
-import random
+import onnxruntime as ort
+import torchvision.transforms as transforms
 import json
 import os
+import numpy as np
+from huggingface_hub import snapshot_download, HfApi
+from transformers import CLIPTokenizer
+
+# --- Config ---
+HUB_REPO_ID = "aarodi/OpenArenaLeaderboard"
+HF_TOKEN = os.environ.get("HF_TOKEN")
+LOCAL_JSON = "leaderboard.json"
+HUB_JSON = "leaderboard.json"
+MODEL_PATH = "mobilenet_v2_fake_detector.onnx"
+CLIP_IMAGE_ENCODER_PATH = "clip_image_encoder.onnx"
+CLIP_TEXT_ENCODER_PATH = "clip_text_encoder.onnx"
+PROMPT_MATCH_THRESHOLD = 10  # percent
+
+# --- Download leaderboard + model checkpoint from HF Hub ---
+def load_assets():
+    try:
+        snapshot_download(
+            repo_id=HUB_REPO_ID,
+            local_dir=".",
+            repo_type="dataset",
+            token=HF_TOKEN,
+            allow_patterns=[HUB_JSON, MODEL_PATH, CLIP_IMAGE_ENCODER_PATH, CLIP_TEXT_ENCODER_PATH]
+        )
+    except Exception as e:
+        print(f"Failed to load assets from HF Hub: {e}")
 
-LEADERBOARD_FILE = "leaderboard.json"
+load_assets()
 
-# --- Load leaderboard from file if it exists ---
-if os.path.exists(LEADERBOARD_FILE):
-    with open(LEADERBOARD_FILE, "r") as f:
-        leaderboard_scores = json.load(f)
-else:
-    leaderboard_scores = {}
+# --- Load leaderboard ---
+def load_leaderboard():
+    try:
+        with open(HUB_JSON, "r") as f:
+            return json.load(f)
+    except Exception as e:
+        print(f"Failed to read leaderboard: {e}")
+        return {}
 
-# --- Save leaderboard to file ---
-def save_leaderboard():
-    with open(LEADERBOARD_FILE, "w") as f:
-        json.dump(leaderboard_scores, f)
+leaderboard_scores = load_leaderboard()
 
-# --- Main logic for prediction ---
-def dummy_deepfake_detector(image: Image.Image, prompt: str, username: str):
+# --- Save and push to HF Hub ---
+def save_leaderboard():
+    try:
+        with open(HUB_JSON, "w") as f:
+            json.dump(leaderboard_scores, f)
+
+        if HF_TOKEN is None:
+            print("HF_TOKEN not set. Skipping push to hub.")
+            return
+
+        api = HfApi()
+        api.upload_file(
+            path_or_fileobj=HUB_JSON,
+            path_in_repo=HUB_JSON,
+            repo_id=HUB_REPO_ID,
+            repo_type="dataset",
+            token=HF_TOKEN,
+            commit_message="Update leaderboard"
+        )
+    except Exception as e:
+        print(f"Failed to save leaderboard to HF Hub: {e}")
+
+# --- Load ONNX models ---
+session = ort.InferenceSession(MODEL_PATH, providers=["CPUExecutionProvider"])
+input_name = session.get_inputs()[0].name
+
+clip_image_sess = ort.InferenceSession(CLIP_IMAGE_ENCODER_PATH, providers=["CPUExecutionProvider"])
+clip_text_sess = ort.InferenceSession(CLIP_TEXT_ENCODER_PATH, providers=["CPUExecutionProvider"])
+clip_tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+transform = transforms.Compose([
+    transforms.Resize((224, 224)),
+    transforms.ToTensor(),
+    transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711])
+])
+
+def compute_prompt_match(image: Image.Image, prompt: str) -> float:
+    try:
+        # Encode image
+        img_tensor = transform(image).unsqueeze(0).numpy().astype(np.float32)
+        image_features = clip_image_sess.run(None, {clip_image_sess.get_inputs()[0].name: img_tensor})[0][0]
+        image_features /= np.linalg.norm(image_features)  # Normalize
+
+        # Encode text
+        inputs = clip_tokenizer(prompt, return_tensors="np", padding="max_length", truncation=True, max_length=77)
+        input_ids = inputs["input_ids"]
+        attention_mask = inputs["attention_mask"]
+        text_features = clip_text_sess.run(None, {
+            clip_text_sess.get_inputs()[0].name: input_ids,
+            clip_text_sess.get_inputs()[1].name: attention_mask
+        })[0][0]
+        text_features /= np.linalg.norm(text_features)  # Normalize
+
+        # Cosine similarity
+        sim = np.dot(image_features, text_features)
+        return round(sim * 100, 2)
+    except Exception as e:
+        print(f"CLIP ONNX match failed: {e}")
+        return 0.0
+
+
+# --- Main prediction logic ---
+def detect_with_model(image: Image.Image, prompt: str, username: str):
     if not username.strip():
-        return "Please enter your name.", None, [], gr.update(visible=True), gr.update(visible=True), gr.update(visible=True), gr.update(visible=False)
+        return "Please enter your name.", None, [], gr.update(visible=True), gr.update(visible=False)
+
+    prompt_score = compute_prompt_match(image, prompt)
+    if prompt_score < PROMPT_MATCH_THRESHOLD:
+        message = f"⚠️ Prompt match too low ({round(prompt_score, 2)}%). Please generate an image that better matches the prompt."
+        return message, None, [], gr.update(visible=True), gr.update(visible=False)
+
+    image_tensor = transforms.Resize((224, 224))(image)
+    image_tensor = transforms.ToTensor()(image_tensor).unsqueeze(0).numpy().astype(np.float32)
+    outputs = session.run(None, {input_name: image_tensor})
+    prob = round(1 / (1 + np.exp(-outputs[0][0][0])), 2)  # sigmoid
+    prediction = "Real" if prob > 0.5 else "Fake"
 
-    prediction = random.choice(["Real", "Fake"])
     score = 1 if prediction == "Real" else 0
+    confidence = round(prob * 100, 2) if prediction == "Real" else round((1 - prob) * 100, 2)
+
+    message = f"Prediction: {prediction} ({confidence}% confidence)\n🧐 Prompt match: {prompt_score}%"
+
+    if prediction == "Real":
+        leaderboard_scores[username] = leaderboard_scores.get(username, 0) + score
+        message += "\n🎉 Nice! You fooled the AI. +1 point!"
+    else:
+        message += "\n😅 The AI caught you this time. Try again!"
 
-    # Update and persist leaderboard
-    leaderboard_scores[username] = leaderboard_scores.get(username, 0) + score
     save_leaderboard()
 
-    # Create sorted leaderboard table
     sorted_scores = sorted(leaderboard_scores.items(), key=lambda x: x[1], reverse=True)
     leaderboard_table = [[name, points] for name, points in sorted_scores]
 
     return (
-        f"Prediction: {prediction}",
+        message,
         image,
         leaderboard_table,
-        gr.update(visible=False),  # hide image input
-        gr.update(visible=False),  # hide prompt input
-        gr.update(visible=False),  # hide upload button
-        gr.update(visible=True)    # show Try Again button
-    )
-
-# --- Reset app state ---
-def reset_app():
-    return (
-        "",                       # Clear prediction text
-        None,                    # Clear image output
-        [],                      # Clear leaderboard
-        gr.update(visible=True, value=""),     # Show prompt input
-        gr.update(visible=True, value=None),   # Show image input
-        gr.update(visible=True),               # Show Upload button
-        gr.update(visible=False),              # Hide Try Again button
-        gr.update(visible=True, value="")      # Show username input
+        gr.update(visible=False),
+        gr.update(visible=True)
     )
 
-# --- Build Gradio UI ---
+# --- UI Layout ---
 with gr.Blocks(css=".gr-button {font-size: 16px !important}") as demo:
-    gr.Markdown("## 🕵️‍♂️ Fool the Deepfake Detector")
-    gr.Markdown("Upload an image and try to fool the AI model into thinking it’s real. Your score will be saved!")
+    gr.Markdown("## 🎝 OpenFake Arena")
+    gr.Markdown("Welcome to the OpenFake Arena!\n\n**Your mission:** Generate a synthetic image for the prompt, upload it, and try to fool the AI detector into thinking it’s real.\n\n**Rules:**\n- Only synthetic images allowed!\n- No cheating with real photos.\n- Licensing is your responsibility.\n\nMake it wild. Make it weird. Most of all — make it fun.")
 
-    with gr.Group():
+    with gr.Group(visible=True) as input_section:
         username_input = gr.Textbox(label="Your Name", placeholder="Enter your name")
 
         with gr.Row():
@@ -74,7 +162,7 @@ with gr.Blocks(css=".gr-button {font-size: 16px !important}") as demo:
             )
 
         with gr.Row():
-            image_input = gr.Image(type="pil", label="Upload Image")
+            image_input = gr.Image(type="pil", label="Upload Synthetic Image")
 
         with gr.Row():
             submit_btn = gr.Button("Upload")
@@ -96,36 +184,28 @@ with gr.Blocks(css=".gr-button {font-size: 16px !important}") as demo:
             row_count=5
         )
 
-    # Submit button logic
     submit_btn.click(
-        fn=dummy_deepfake_detector,
+        fn=detect_with_model,
         inputs=[image_input, prompt_input, username_input],
         outputs=[
             prediction_output,
             image_output,
             leaderboard,
-            image_input,
-            prompt_input,
-            submit_btn,
+            input_section,
             try_again_btn
         ]
     )
 
-    # Try Again button logic
     try_again_btn.click(
-        fn=reset_app,
+        fn=lambda: ("", None, [], gr.update(visible=True), gr.update(visible=False)),
         outputs=[
             prediction_output,
             image_output,
             leaderboard,
-            prompt_input,
-            image_input,
-            submit_btn,
-            try_again_btn,
-            username_input
+            input_section,
+            try_again_btn
         ]
     )
 
 if __name__ == "__main__":
     demo.launch()
-

requirements.txt

@@ -1,2 +1,4 @@
 gradio
 pillow
+onnxruntime
+scikit-image