Upload app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,152 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import spaces
|
| 2 |
+
import gradio as gr
|
| 3 |
+
import torch
|
| 4 |
+
from transformers import CLIPModel, CLIPProcessor
|
| 5 |
+
from PIL import Image
|
| 6 |
+
|
| 7 |
+
# ─── Model Loading (module-level for ZeroGPU) ───────────────────────────
|
| 8 |
+
MODEL_ID = "kwanY/styleid"
|
| 9 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 10 |
+
model = CLIPModel.from_pretrained(MODEL_ID).to(device)
|
| 11 |
+
processor = CLIPProcessor.from_pretrained(MODEL_ID)
|
| 12 |
+
model.eval()
|
| 13 |
+
|
| 14 |
+
# ─── Inference ───────────────────────────────────────────────────────────
|
| 15 |
+
|
| 16 |
+
@spaces.GPU(duration=30)
|
| 17 |
+
def compare_faces(img1: Image.Image, img2: Image.Image):
|
| 18 |
+
"""Compare face identity between two images using StyleID."""
|
| 19 |
+
if img1 is None or img2 is None:
|
| 20 |
+
return "⚠️ Please upload both images.", ""
|
| 21 |
+
|
| 22 |
+
def get_embedding(img):
|
| 23 |
+
inputs = processor(images=img.convert("RGB"), return_tensors="pt").to(device)
|
| 24 |
+
with torch.no_grad():
|
| 25 |
+
emb = model.get_image_features(**inputs)
|
| 26 |
+
emb = emb / emb.norm(dim=-1, keepdim=True)
|
| 27 |
+
return emb
|
| 28 |
+
|
| 29 |
+
e1 = get_embedding(img1)
|
| 30 |
+
e2 = get_embedding(img2)
|
| 31 |
+
score = (e1 * e2).sum(dim=-1).item()
|
| 32 |
+
|
| 33 |
+
# Interpretation based on paper thresholds
|
| 34 |
+
if score >= 0.55:
|
| 35 |
+
verdict = "✅ Confidently Same Person"
|
| 36 |
+
color = "#22c55e"
|
| 37 |
+
elif score >= 0.45:
|
| 38 |
+
verdict = "✅ Very Likely Same Person"
|
| 39 |
+
color = "#4ade80"
|
| 40 |
+
elif score >= 0.35:
|
| 41 |
+
verdict = "🟡 Probably Same Person"
|
| 42 |
+
color = "#facc15"
|
| 43 |
+
elif score >= 0.25:
|
| 44 |
+
verdict = "🟠 Uncertain — Borderline"
|
| 45 |
+
color = "#fb923c"
|
| 46 |
+
else:
|
| 47 |
+
verdict = "❌ Likely Different People"
|
| 48 |
+
color = "#ef4444"
|
| 49 |
+
|
| 50 |
+
# Create a visual gauge
|
| 51 |
+
score_pct = max(0, min(100, int(score * 100)))
|
| 52 |
+
gauge_html = f"""
|
| 53 |
+
<div style="text-align:center; padding: 20px;">
|
| 54 |
+
<div style="font-size: 48px; font-weight: bold; color: {color};">{score:.3f}</div>
|
| 55 |
+
<div style="font-size: 20px; margin-top: 8px; color: {color}; font-weight: 600;">{verdict}</div>
|
| 56 |
+
<div style="margin-top: 16px; background: #1f2937; border-radius: 12px; height: 24px; overflow: hidden; position: relative;">
|
| 57 |
+
<div style="background: linear-gradient(90deg, #ef4444, #fb923c, #facc15, #4ade80, #22c55e); height: 100%; width: {score_pct}%; border-radius: 12px; transition: width 0.5s;"></div>
|
| 58 |
+
</div>
|
| 59 |
+
<div style="display: flex; justify-content: space-between; font-size: 11px; color: #9ca3af; margin-top: 4px;">
|
| 60 |
+
<span>Different</span>
|
| 61 |
+
<span>Uncertain</span>
|
| 62 |
+
<span>Same</span>
|
| 63 |
+
</div>
|
| 64 |
+
</div>
|
| 65 |
+
"""
|
| 66 |
+
|
| 67 |
+
details = f"""### How to Read the Score
|
| 68 |
+
| Range | Meaning |
|
| 69 |
+
|-------|---------|
|
| 70 |
+
| **> 0.55** | ✅ Confidently same person |
|
| 71 |
+
| **0.45 – 0.55** | ✅ Very likely same person |
|
| 72 |
+
| **0.35 – 0.45** | 🟡 Probably same person |
|
| 73 |
+
| **0.25 – 0.35** | 🟠 Uncertain / borderline |
|
| 74 |
+
| **< 0.25** | ❌ Likely different people |
|
| 75 |
+
|
| 76 |
+
**Your Score: {score:.4f}**
|
| 77 |
+
|
| 78 |
+
> **Note:** StyleID is optimized for comparing a real photo against its stylized version (anime, cartoon, painting, sketch, etc.). Works best with a single prominent face per image.
|
| 79 |
+
"""
|
| 80 |
+
|
| 81 |
+
return gauge_html, details
|
| 82 |
+
|
| 83 |
+
# ─── UI ──────────────────────────────────────────────────────────────────
|
| 84 |
+
|
| 85 |
+
DESCRIPTION = """
|
| 86 |
+
# 🎭 StyleID — Does AI Preserve Your Face?
|
| 87 |
+
|
| 88 |
+
Upload a **real photo** and a **stylized portrait** (anime, cartoon, painting, sketch, etc.)
|
| 89 |
+
→ StyleID checks if the identity is preserved across art styles.
|
| 90 |
+
|
| 91 |
+
**Powered by [StyleID](https://huggingface.co/kwanY/styleid)** — a perception-aware face identity model trained to match human judgments across stylizations.
|
| 92 |
+
|
| 93 |
+
### 🔬 How it works
|
| 94 |
+
StyleID uses a fine-tuned CLIP ViT-L/14 encoder calibrated against human psychometric experiments.
|
| 95 |
+
Unlike standard face recognition (ArcFace, AdaFace), StyleID is specifically designed for **cross-style** identity matching —
|
| 96 |
+
real photo ↔ anime, cartoon, painting, caricature, 3D render, etc.
|
| 97 |
+
|
| 98 |
+
### ⚡ Use Cases
|
| 99 |
+
- Check if your AI-stylized portrait still looks like you
|
| 100 |
+
- Evaluate face-swap / face-transfer quality
|
| 101 |
+
- Compare identity preservation across different AI art tools
|
| 102 |
+
- Research: benchmark stylization models for ID preservation
|
| 103 |
+
"""
|
| 104 |
+
|
| 105 |
+
TIPS = """
|
| 106 |
+
### 💡 Tips for Best Results
|
| 107 |
+
1. **One face per image** — StyleID doesn't do face detection
|
| 108 |
+
2. **Face should be prominent** — at least 25% of the image
|
| 109 |
+
3. **Frontal or 3/4 view works best** — extreme profiles may reduce scores
|
| 110 |
+
4. **Works across styles** — anime, cartoon, oil painting, watercolor, pencil sketch, 3D render, caricature
|
| 111 |
+
5. **NOT a security tool** — for research/evaluation only
|
| 112 |
+
"""
|
| 113 |
+
|
| 114 |
+
with gr.Blocks(
|
| 115 |
+
title="StyleID — Face Identity Across Art Styles",
|
| 116 |
+
theme=gr.themes.Soft(
|
| 117 |
+
primary_hue="violet",
|
| 118 |
+
secondary_hue="blue",
|
| 119 |
+
),
|
| 120 |
+
css="footer { display: none !important; }"
|
| 121 |
+
) as demo:
|
| 122 |
+
gr.Markdown(DESCRIPTION)
|
| 123 |
+
|
| 124 |
+
with gr.Row():
|
| 125 |
+
with gr.Column():
|
| 126 |
+
img1 = gr.Image(type="pil", label="📸 Original Photo (Real Face)", height=400)
|
| 127 |
+
with gr.Column():
|
| 128 |
+
img2 = gr.Image(type="pil", label="🎨 Stylized Portrait (Anime, Cartoon, etc.)", height=400)
|
| 129 |
+
|
| 130 |
+
btn = gr.Button("🔍 Compare Identity", variant="primary", size="lg")
|
| 131 |
+
|
| 132 |
+
with gr.Row():
|
| 133 |
+
with gr.Column():
|
| 134 |
+
gauge = gr.HTML(label="Identity Score")
|
| 135 |
+
with gr.Column():
|
| 136 |
+
details = gr.Markdown(label="Details")
|
| 137 |
+
|
| 138 |
+
btn.click(
|
| 139 |
+
fn=compare_faces,
|
| 140 |
+
inputs=[img1, img2],
|
| 141 |
+
outputs=[gauge, details],
|
| 142 |
+
)
|
| 143 |
+
|
| 144 |
+
with gr.Accordion("💡 Tips & Info", open=False):
|
| 145 |
+
gr.Markdown(TIPS)
|
| 146 |
+
|
| 147 |
+
gr.Markdown("""
|
| 148 |
+
---
|
| 149 |
+
**Model:** [kwanY/styleid](https://huggingface.co/kwanY/styleid) | **Paper:** [StyleID (arXiv 2604.21689)](https://arxiv.org/abs/2604.21689) | **License:** Research use only — not for biometric authentication.
|
| 150 |
+
""")
|
| 151 |
+
|
| 152 |
+
demo.launch()
|