Upload 9 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+prompts_0.csv filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,12 +1,13 @@
 ---
-title: OpenFakeArena
-emoji: 💻
-colorFrom: pink
-colorTo: gray
+title: Fool The AI Detector
+emoji: 🌍
+colorFrom: red
+colorTo: indigo
 sdk: gradio
-sdk_version: 6.0.0
+sdk_version: 5.29.0
 app_file: app.py
 pinned: false
+short_description: Fooling an synthetic image detector
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+This is the OpenFake Arena.

app.py

@@ -0,0 +1,282 @@
+import gradio as gr
+from PIL import Image
+import torch
+import torchvision.transforms as transforms
+import json
+import os
+import numpy as np
+import pandas as pd
+import random
+import onnxruntime as ort
+from transformers import CLIPTokenizer, AutoImageProcessor, AutoModelForImageClassification
+from safetensors.torch import load_file as safe_load
+from datetime import datetime
+
+# --- Config ---
+LEADERBOARD_JSON = "leaderboard.json"
+MODEL_PATH = "model.safetensors"  # ✅ updated filename
+MODEL_BACKBONE = "microsoft/swinv2-small-patch4-window16-256"
+CLIP_IMAGE_ENCODER_PATH = "clip_image_encoder.onnx"
+CLIP_TEXT_ENCODER_PATH = "clip_text_encoder.onnx"
+PROMPT_CSV_PATH = "prompts_0.csv"
+PROMPT_MATCH_THRESHOLD = 25  # percent
+
+# --- No-op for HF Space ---
+def load_assets():
+    print("Skipping snapshot_download. Assuming files exist via Git LFS in HF Space.")
+
+load_assets()
+
+# --- Load leaderboard ---
+def load_leaderboard():
+    try:
+        with open(LEADERBOARD_JSON, "r", encoding="utf-8") as f:
+            return json.load(f)
+    except Exception as e:
+        print(f"Failed to load leaderboard: {e}")
+        return {}
+
+leaderboard_scores = load_leaderboard()
+
+def save_leaderboard():
+    try:
+        with open(LEADERBOARD_JSON, "w", encoding="utf-8") as f:
+            json.dump(leaderboard_scores, f, ensure_ascii=False)
+    except Exception as e:
+        print(f"Failed to save leaderboard: {e}")
+
+# --- Load prompts from CSV ---
+def load_prompts():
+    try:
+        df = pd.read_csv(PROMPT_CSV_PATH)
+        if "prompt" in df.columns:
+            return df["prompt"].dropna().tolist()
+        else:
+            print("CSV missing 'prompt' column.")
+            return []
+    except Exception as e:
+        print(f"Failed to load prompts: {e}")
+        return []
+
+PROMPT_LIST = load_prompts()
+
+# --- Load model + processor ---
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+processor = AutoImageProcessor.from_pretrained(MODEL_BACKBONE)
+model = AutoModelForImageClassification.from_pretrained(MODEL_BACKBONE)
+model.classifier = torch.nn.Linear(model.config.hidden_size, 2)
+
+model.load_state_dict(safe_load(MODEL_PATH, device="cpu"), strict=False)
+model.to(device)
+model.eval()
+
+# --- CLIP prompt matching ---
+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((256, 256)),
+    transforms.ToTensor(),
+    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+])
+
+def compute_prompt_match(image: Image.Image, prompt: str) -> float:
+    try:
+        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)
+
+        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)
+
+        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, model_name: str):
+    if not username.strip():
+        return "Please enter your name.", None, [], gr.update(visible=True), gr.update(visible=False), username
+
+    prompt_score = compute_prompt_match(image, prompt)
+    if prompt_score < PROMPT_MATCH_THRESHOLD and (model_name.lower() != "real" or model_name != ""):
+        message = f"⚠️ Prompt match too low ({round(prompt_score, 2)}%). Please generate an image that better matches the prompt."
+        return message, None, leaderboard, gr.update(visible=True), gr.update(visible=False), username
+
+    # Run model inference
+    inputs = processor(image, return_tensors="pt").to(device)
+    with torch.no_grad():
+        outputs = model(**inputs)
+        logits = outputs.logits
+        pred_class = torch.argmax(logits, dim=-1).item()
+        prediction = "Real" if pred_class == 0 else "Fake"
+
+        probs = torch.softmax(logits, dim=-1)[0]
+        confidence = round(probs[pred_class].item() * 100, 2)
+
+    score = 1 if prediction == "Real" else 0
+
+    message = f"🔍 Prediction: {prediction} ({confidence}% confidence)\n🧐 Prompt match: {round(prompt_score, 2)}%"
+    if prediction == "Real" and model_name.lower() != "real":
+        leaderboard_scores[username] = leaderboard_scores.get(username, 0) + score
+        message += "\n🎉 Nice! You fooled the AI. +1 point!"            
+    else:
+        if model_name.lower() == "real":
+            message += "\n You uploaded a real image, this does not count toward the leaderboard!"
+        else:
+            message += "\n😅 The AI caught you this time. Try again!"
+
+    save_leaderboard()
+
+    sorted_scores = sorted(leaderboard_scores.items(), key=lambda x: x[1], reverse=True)
+    leaderboard_table = [[name, points] for name, points in sorted_scores]
+
+    image_path = None
+    try:
+        type_image = "real" if (model_name.lower() == "real" or model_name == "") else "fake"
+        image_dir = os.path.join("test", type_image)
+        os.makedirs(image_dir, exist_ok=True)
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        image_filename = f"{timestamp}.jpg"
+        image_path = os.path.join(image_dir, image_filename)
+        image.save(image_path)
+    except Exception as e:
+        print(f"Failed to save image locally: {e}")
+    finally:
+        if image_path and os.path.exists(image_path):
+            try:
+                os.remove(image_path)
+            except Exception as cleanup_error:
+                print(f"Failed to delete temporary image: {cleanup_error}")
+
+    return (
+        message,
+        image,
+        leaderboard_table,
+        gr.update(visible=False),
+        gr.update(visible=True),
+        username
+    )
+
+def get_random_prompt():
+    return random.choice(PROMPT_LIST) if PROMPT_LIST else "A synthetic scene with dramatic lighting"
+
+def load_initial_state():
+    sorted_scores = sorted(leaderboard_scores.items(), key=lambda x: x[1], reverse=True)
+    leaderboard_table = [[name, points] for name, points in sorted_scores]
+    return gr.update(value=get_random_prompt()), leaderboard_table
+
+# --- Gradio UI ---
+with gr.Blocks(css=".gr-button {font-size: 16px !important}") as demo:
+    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\n- You can modify the prompt on your end, but the image needs to have the same content. We verify the content with a CLIP similarity threshold.\n\n- Enter \"real\" in the model used to upload and test a real image. You don't need to follow the prompt for real images. Tips: you can also enter \"real\" if you just want to test the detector! We won't be collecting those images. \n\n- It is important to enter the correct model name for licensing.\n\n- Only synthetic images count toward the leaderboard!\n\n\nNote: The detector is still in early development. The prompt is not used for prediction, only the image.")
+
+    with gr.Group(visible=True) as input_section:
+        username_input = gr.Textbox(label="Your Name", placeholder="Enter your name", interactive=True)
+        model_input = gr.Textbox(label="Model used, specify the version (e.g., Imagen 3, Dall-e 3,  Midjourney 6). Write \"Real\" when uploading a real image.", placeholder="Name of the model used to generate the image", interactive=True)
+
+        # 🚫 Freeze this block: do not allow edits to the prompt input component's configuration.
+        with gr.Row():
+            prompt_input = gr.Textbox(
+                interactive=False,
+                label="Prompt to match",
+                placeholder="e.g., ...",
+                value="",
+                lines=2
+            )
+
+        with gr.Row():
+            image_input = gr.Image(type="pil", label="Upload Synthetic Image")
+
+        with gr.Row():
+            submit_btn = gr.Button("Upload")
+
+    try_again_btn = gr.Button("Try Again", visible=False)
+
+    with gr.Group():
+        gr.Markdown("### 🎯 Result")
+        with gr.Row():
+            prediction_output = gr.Textbox(label="Prediction", interactive=False, elem_id="prediction_box")
+            image_output = gr.Image(label="Submitted Image", show_label=False)
+
+    with gr.Group():
+        gr.Markdown("### 🏆 Leaderboard")
+        leaderboard = gr.Dataframe(
+            headers=["Username", "Score"],
+            datatype=["str", "number"],
+            interactive=False,
+            row_count=5,
+            visible=True
+        )
+
+    submit_btn.click(
+        fn=detect_with_model,
+        inputs=[image_input, prompt_input, username_input, model_input],
+        outputs=[
+            prediction_output,
+            image_output,
+            leaderboard,
+            input_section,
+            try_again_btn,
+            username_input
+        ]
+    )
+
+    try_again_btn.click(
+        fn=lambda name: (
+            "",               # Clear prediction text
+            None,             # Clear uploaded image
+            leaderboard,               # Clear leaderboard (temporarily, gets reloaded on next submit)
+            gr.update(visible=True),   # Show input section
+            gr.update(visible=False),  # Hide "Try Again" button
+            name,             # Keep username
+            gr.update(value=get_random_prompt()),  # Load new prompt
+            None              # Clear image input
+        ),
+        inputs=[username_input],
+        outputs=[
+            prediction_output,
+            image_output,
+            leaderboard,
+            input_section,
+            try_again_btn,
+            username_input,
+            prompt_input,
+            image_input        # ← added output to clear image
+        ]
+    )
+
+    demo.load(
+        fn=load_initial_state,
+        outputs=[prompt_input, leaderboard]
+    )
+
+
+    gr.HTML("""
+    <script>
+    document.addEventListener('DOMContentLoaded', function () {
+        const target = document.getElementById('prediction_box');
+        const observer = new MutationObserver(() => {
+            if (target && target.innerText.trim() !== '') {
+                window.scrollTo({ top: 0, behavior: 'smooth' });
+            }
+        });
+        if (target) {
+            observer.observe(target, { childList: true, subtree: true });
+        }
+    });
+    </script>
+    """)
+
+if __name__ == "__main__":
+    demo.launch()

clip_image_encoder.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e557b21fad17709c7a0f4133ad24206cc887f551e58683c200d4afecaba2289c
+size 351635998

clip_text_encoder.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66bab8550f6f98ad825b5df3cf07f31c4d61e2e15dae4c9bc6231949c8faba51
+size 253958912

leaderboard.json

@@ -0,0 +1 @@
+{"Anonymous": 1}

mobilenet_v2_fake_detector.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fd71df3f186e80b862a4d367ef19e2d19b29ae70588092dce16bfb92e7abbe3
+size 8878521

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fdae478dc277ab1284a8e06b7934334ed8277d21371a8c95a8abfcd67368631
+size 195905448

prompts_0.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:123f0b62272dba0ecc69b565a09df93893242cbe2a41f5171858a281f7c6cd40
+size 45613056

requirements.txt

@@ -0,0 +1,7 @@
+gradio
+pillow
+onnxruntime
+scikit-image
+torchvision
+torch
+transformers