README.md

metadata

title: FaceSwap GPU
emoji: 🎬
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.29.0
app_file: app.py
pinned: false
license: unknown
short_description: GPU-accelerated all-in-one face swapping (photo + video)

FaceSwap GPU – Rented-GPU Ready

All-in-one face swapping suite optimised for CUDA GPUs (RunPod, Vast.ai, Lambda, local RTX, HF GPU Spaces, etc.).

What changed vs the original

• Explicit CUDAExecutionProvider preference via onnxruntime
• Robust Gradio video path handling (fixes all video tabs)
• Cleaned frame extraction + resume logic
• Better error reporting in the Log box
• Workdir isolation under workdir/
• opencv-python-headless (safer on servers)
• Proper demo.queue() for concurrent jobs
• ssr_mode=False to avoid Gradio SSR 405 errors
• ffmpeg system package declared

Requirements on the rented machine

• NVIDIA GPU + drivers
• CUDA-compatible onnxruntime-gpu (already in requirements)
• ~4–8 GB VRAM recommended for 1080p video
• ffmpeg installed (packages.txt)

Local / rented GPU launch

pip install -r requirements.txt
python app.py --share          # creates a public link
# or
python app.py --server-port 7860

First run downloads buffalo_l + inswapper_128.onnx (~300 MB).

Tips for long videos

• Keep “Delete extracted frames” checked to save disk.
• Processing is frame-by-frame; a 1-minute 30 fps clip ≈ 1800 swaps.
• On a modern GPU expect 5–15 fps effective (depends on resolution & number of faces).
• Resume is automatic: if the process is killed, re-run and already-swapped frames are skipped.

Face indices

Faces are sorted left-to-right. Index 1 = leftmost face.