import os
import urllib.request
import numpy as np
import cv2
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms as transforms
import timm
import gradio as gr
from PIL import Image
# =====================
# CONFIG
# =====================
MODEL_URL = "https://huggingface.co/ARPAN2026/dfake-hcnext/resolve/main/best_model_New.pth"
MODEL_PATH = "best_model_New.pth"
IMG_SIZE = 224
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# =====================
# MODEL DEFINITION
# =====================
class DeepfakeModel(nn.Module):
def __init__(self):
super().__init__()
self.backbone = timm.create_model("convnext_base", pretrained=False, num_classes=0)
dim = self.backbone.num_features
self.classifier = nn.Sequential(
nn.LayerNorm(dim),
nn.Linear(dim, 256),
nn.GELU(),
nn.Dropout(0.4),
nn.Linear(256, 2),
)
def forward(self, x):
f = self.backbone.forward_features(x)
if len(f.shape) == 4:
f = f.flatten(2).mean(-1)
return self.classifier(f)
# =====================
# DOWNLOAD + LOAD
# =====================
def download_model():
if not os.path.exists(MODEL_PATH):
print("Downloading model weights…")
urllib.request.urlretrieve(MODEL_URL, MODEL_PATH)
print("Download complete.")
download_model()
model = DeepfakeModel().to(DEVICE)
model.load_state_dict(torch.load(MODEL_PATH, map_location=DEVICE))
model.eval()
print("Model loaded successfully.")
# =====================
# TRANSFORM
# =====================
transform = transforms.Compose([
transforms.Resize((IMG_SIZE, IMG_SIZE)),
transforms.ToTensor(),
transforms.Normalize([0.5] * 3, [0.5] * 3),
])
# =====================
# INFERENCE
# =====================
def predict(image: Image.Image):
if image is None:
return {"Error": 1.0}, "⚠️ Please upload an image."
img_tensor = transform(image.convert("RGB")).unsqueeze(0).to(DEVICE)
with torch.no_grad():
logits = model(img_tensor)
probs = torch.softmax(logits, dim=1).cpu().numpy()[0]
real_prob = float(probs[0])
fake_prob = float(probs[1])
confidence = max(real_prob, fake_prob) * 100
if fake_prob > real_prob:
verdict = "🔴 DEEPFAKE DETECTED"
verdict_md = f"## {verdict}\n**Confidence:** {confidence:.1f}%"
else:
verdict = "🟢 LIKELY REAL"
verdict_md = f"## {verdict}\n**Confidence:** {confidence:.1f}%"
label_dict = {
"Real": round(real_prob, 4),
"Fake": round(fake_prob, 4),
}
return label_dict, verdict_md
# =====================
# CUSTOM CSS (dark forensic theme)
# =====================
CSS = """
@import url('https://fonts.googleapis.com/css2?family=Share+Tech+Mono&family=Syne:wght@400;700;800&display=swap');
:root {
--bg: #0a0c10;
--surface: #111318;
--border: #1e2330;
--accent: #00e5ff;
--danger: #ff3b5c;
--safe: #00e676;
--text: #d0d8f0;
--muted: #5a6480;
--radius: 8px;
}
body, .gradio-container {
background: var(--bg) !important;
font-family: 'Syne', sans-serif !important;
color: var(--text) !important;
}
/* ---- header ---- */
.gr-header {
background: transparent !important;
}
h1.title-heading {
font-family: 'Syne', sans-serif;
font-weight: 800;
font-size: 2.4rem;
letter-spacing: -0.02em;
background: linear-gradient(90deg, var(--accent), #7b61ff);
-webkit-background-clip: text;
-webkit-text-fill-color: transparent;
margin: 0;
}
p.subtitle {
color: var(--muted);
font-family: 'Share Tech Mono', monospace;
font-size: 0.85rem;
margin-top: 4px;
letter-spacing: 0.08em;
}
/* ---- panels ---- */
.gr-box, .gr-panel, .gr-form {
background: var(--surface) !important;
border: 1px solid var(--border) !important;
border-radius: var(--radius) !important;
}
/* ---- upload zone ---- */
.gr-image, .svelte-1n8nu59 {
border: 2px dashed var(--border) !important;
border-radius: var(--radius) !important;
background: #0d0f14 !important;
}
/* ---- buttons ---- */
button.primary {
background: var(--accent) !important;
color: #000 !important;
font-family: 'Syne', sans-serif !important;
font-weight: 700 !important;
border: none !important;
border-radius: var(--radius) !important;
letter-spacing: 0.05em;
}
button.secondary {
background: transparent !important;
border: 1px solid var(--border) !important;
color: var(--muted) !important;
font-family: 'Syne', sans-serif !important;
border-radius: var(--radius) !important;
}
/* ---- labels / markdown output ---- */
.gr-markdown h2 {
font-family: 'Syne', sans-serif;
font-size: 1.4rem;
font-weight: 700;
margin: 0 0 4px;
}
/* ---- confidence bars ---- */
.gr-label .wrap {
background: var(--surface) !important;
border: 1px solid var(--border) !important;
border-radius: var(--radius) !important;
}
.gr-label .label-wrap span {
font-family: 'Share Tech Mono', monospace !important;
color: var(--text) !important;
}
/* confidence fill colors */
.gr-label .bar {
background: linear-gradient(90deg, var(--accent), #7b61ff) !important;
}
/* ---- footer ---- */
footer { display: none !important; }
"""
# =====================
# GRADIO UI
# =====================
with gr.Blocks(css=CSS, title="DeepFake Detector") as demo:
gr.HTML("""
DEEPFAKE DETECTOR
ConvNeXt-Base · Trained on RVF Faces · Hackathon Edition
""")
with gr.Row():
with gr.Column(scale=1):
image_input = gr.Image(
type="pil",
label="Upload Face Image",
height=320,
)
with gr.Row():
submit_btn = gr.Button("🔍 Analyze", variant="primary")
clear_btn = gr.ClearButton([image_input], value="✕ Clear")
gr.HTML("""
MODEL ConvNeXt-Base + custom head
TRAINED Real vs Fake Faces (80/20 split)
INPUT 224 × 224 · RGB · normalized
CLASSES Real · Fake
""")
with gr.Column(scale=1):
verdict_output = gr.Markdown(
value="*Upload an image and click **Analyze** to begin.*",
label="Verdict",
)
label_output = gr.Label(
num_top_classes=2,
label="Class Probabilities",
)
# example images (optional — works if you add them to the Space repo)
gr.Examples(
examples=[], # add paths like [["examples/real1.jpg"], ["examples/fake1.jpg"]]
inputs=image_input,
label="Example Images",
)
submit_btn.click(
fn=predict,
inputs=image_input,
outputs=[label_output, verdict_output],
)
gr.HTML("""
Built with ❤ · Gradio · HuggingFace Spaces · PyTorch
""")
demo.launch()