README.md

metadata

title: Video Face Replace
emoji: 🎭
colorFrom: purple
colorTo: pink
sdk: gradio
sdk_version: 6.10.0
app_file: app.py
pinned: true
license: mit
short_description: Swap any face into any video using InsightFace + GFPGAN

🎭 Face Fusion — AI Video Face Swap

Swap any face into any video using state-of-the-art AI. Built on InsightFace, inswapper_128, and GFPGAN.

⚡ This is a Kaggle notebook project — run it there for full GPU-accelerated face swapping. The Gradio UI is embedded inside the notebook (Cell 4).

---

✨ What It Does

Upload a single clear photo of a face, point it at a YouTube video (or upload your own), and the pipeline replaces every detected face in the video with your source face — enhanced, sharpened, and re-merged with the original audio.

Pipeline:

Source Face Photo → InsightFace Embedding
       ↓
Target Video (YouTube URL or upload)
       ↓
Per-frame: Face Detection → Swap (inswapper_128) → GFPGAN Enhancement
       ↓
Audio Merge (ffmpeg) → Final MP4

---

🚀 How to Run (Kaggle)

1. Open Face_replace.ipynb on Kaggle with GPU accelerator (T4 or P100)
2. Cell 1 — installs packages + downloads models (~1 min)
3. Cell 2 — set source face path + YouTube URL, downloads & trims video
4. Cell 3 — processes all frames → outputs face_swapped_output.mp4
5. Cell 4 — launches interactive Gradio UI

Processes ~12 frames/second on T4 GPU

---

🧰 Tech Stack

Component	Role
InsightFace buffalo_l	Face detection + recognition
inswapper_128.onnx	Neural face-swap model
GFPGAN v1.4	Face restoration & enhancement
OpenCV	Frame-level video I/O
ffmpeg	Video codec + audio merge
yt-dlp	YouTube video download
Gradio	Interactive UI
ONNX Runtime GPU	GPU-accelerated inference

---

💡 Tips for Best Results

• Use a clear, front-facing photo — no sunglasses or heavy shadows
• Keep target videos under 30 seconds for fast iteration
• Single-face videos give the cleanest results
• Enable GFPGAN (ENABLE_ENHANCEMENT = True) for sharper output

---

⚠️ Disclaimer

For creative, educational, and research purposes only. Do not use to create deceptive, harmful, or non-consensual content. Always obtain permission before using someone's likeness.

---

🙏 Credits

• deepinsight/insightface
• TencentARC/GFPGAN
• ezioruan/inswapper_128.onnx

---

Made with ❤️ by Aastha Malik