app.py

import os
import cv2
import time
import torch
import argparse
import insightface
import onnxruntime
import numpy as np
import gradio as gr
from tqdm import tqdm

from face_swapper import Inswapper, paste_to_whole
from face_analyser import analyse_face
from face_enhancer import load_face_enhancer_model, cv2_interpolations
from utils import create_image_grid

## ------------------------------ USER ARGS ------------------------------

parser = argparse.ArgumentParser(description="Free Face Swapper (Male/Female mode)")
parser.add_argument("--out_dir", default=os.getcwd())
parser.add_argument("--batch_size", default=32)
parser.add_argument("--cuda", action="store_true", default=False)
user_args = parser.parse_args()

USE_CUDA = user_args.cuda
DEF_OUTPUT_PATH = user_args.out_dir
BATCH_SIZE = int(user_args.batch_size)

## ------------------------------ DEVICE ------------------------------

PROVIDER = ["CPUExecutionProvider"]
if USE_CUDA:
    if "CUDAExecutionProvider" in onnxruntime.get_available_providers():
        PROVIDER = ["CUDAExecutionProvider", "CPUExecutionProvider"]
        print(">>> Running on CUDA")
    else:
        USE_CUDA = False
        print(">>> CUDA not available, running on CPU")

device = "cuda" if USE_CUDA else "cpu"
EMPTY_CACHE = lambda: torch.cuda.empty_cache() if device == "cuda" else None

## ------------------------------ LOAD MODELS ------------------------------

FACE_ANALYSER = insightface.app.FaceAnalysis(name="buffalo_l", providers=PROVIDER)
FACE_ANALYSER.prepare(ctx_id=0, det_size=(640, 640), det_thresh=0.6)

FACE_SWAPPER = Inswapper(
    model_file="./assets/pretrained_models/inswapper_128.onnx",
    batch_size=(BATCH_SIZE if USE_CUDA else 1),
    providers=PROVIDER,
)

## ------------------------------ PROCESS ------------------------------

def swap_faces(image_path, male_source_path, female_source_path, face_enhancer_name="NONE"):
    start_time = time.time()

    # Load target
    target = cv2.imread(image_path)

    # Load source male/female
    analysed_source_male = analyse_face(cv2.imread(male_source_path), FACE_ANALYSER)
    analysed_source_female = analyse_face(cv2.imread(female_source_path), FACE_ANALYSER)

    # Analyse target
    analysed_faces = FACE_ANALYSER.get(target)

    preds, matrs = [], []
    for analysed_face in tqdm(analysed_faces, desc="Swapping faces"):
        if analysed_face["gender"] == 1:  # male
            src = analysed_source_male
        else:  # female
            src = analysed_source_female

        batch_pred, batch_matr = FACE_SWAPPER.get([target], [analysed_face], [src])
        preds.extend(batch_pred)
        matrs.extend(batch_matr)
        EMPTY_CACHE()

    # Paste back
    for p, m in zip(preds, matrs):
        target = paste_to_whole(p, target, m, blend_method="laplacian")

    # Enhance (optional)
    if face_enhancer_name != "NONE":
        model, runner = load_face_enhancer_model(face_enhancer_name, device=device)
        target = runner(target, model)

    elapsed = time.time() - start_time
    print(f"✔ Done in {elapsed:.2f} sec")
    return target[:, :, ::-1]  # BGR->RGB for display


## ------------------------------ GRADIO UI ------------------------------

with gr.Blocks() as demo:
    gr.Markdown("## 🧑➡👩 Face Swapper (Male+Female sources)")

    with gr.Row():
        with gr.Column():
            image_input = gr.Image(label="Target Image", type="filepath")
            male_input = gr.Image(label="Source Male", type="filepath")
            female_input = gr.Image(label="Source Female", type="filepath")
            enhancer = gr.Dropdown(
                ["NONE"] + cv2_interpolations, label="Face Enhancer", value="NONE"
            )
            run_btn = gr.Button("✨ Swap")

        with gr.Column():
            output_image = gr.Image(label="Output")

    run_btn.click(
        fn=swap_faces,
        inputs=[image_input, male_input, female_input, enhancer],
        outputs=output_image,
    )

if __name__ == "__main__":
    demo.launch()