swap.py

import argparse
import os

import cv2
import torch
import numpy as np  # Import numpy explicitly
from PIL import Image  # Use PIL for image processing
from insightface.app import FaceAnalysis
import face_align

faceAnalysis = FaceAnalysis(name='buffalo_l')
faceAnalysis.prepare(ctx_id=-1, det_size=(512, 512)) #ctx_id=-1 for CPU

from StyleTransferModel_128 import StyleTransferModel
import gradio as gr

def parse_arguments():
    parser = argparse.ArgumentParser(description='Process command line arguments')
    
    parser.add_argument('--modelPath', required=True, help='Model path')
    parser.add_argument('--resolution', type=int, default=128, help='Resolution')

    return parser.parse_args()

def get_device():
    return torch.device('cpu')  # Force CPU

def load_model(model_path):
    device = get_device()
    model = StyleTransferModel().to(device)
    model.load_state_dict(torch.load(model_path, map_location=device), strict=False)
    model.eval()
    return model

def swap_face(model, target_face, source_face_latent):
    device = get_device()

    target_tensor = torch.from_numpy(target_face).to(device)
    source_tensor = torch.from_numpy(source_face_latent).to(device)

    with torch.no_grad():
        swapped_tensor = model(target_tensor, source_tensor)
    
    swapped_face = postprocess_face(swapped_tensor) # Use PIL-based postprocess

    return swapped_face, swapped_tensor

def create_target(target_image, resolution):
    target_face = faceAnalysis.get(np.array(target_image))[0] # Convert PIL to numpy

    aligned_target_face, M = face_align.norm_crop2(np.array(target_image), target_face.kps, resolution) # Convert PIL to numpy
    target_face_blob = getBlob(aligned_target_face, (resolution, resolution))

    return target_face_blob, M

def create_source(source_image):
    source_face = faceAnalysis.get(np.array(source_image))[0] # Convert PIL to numpy
    source_latent = getLatent(source_face)

    return source_latent


def postprocess_face(swapped_tensor):
    swapped_tensor = swapped_tensor.cpu().numpy()
    swapped_tensor = np.transpose(swapped_tensor, (0, 2, 3, 1))
    swapped_tensor = (swapped_tensor * 255).astype(np.uint8)
    swapped_face = Image.fromarray(swapped_tensor[0]) # Convert to PIL Image
    return swapped_face

def getBlob(aligned_face, size):
    aligned_face = cv2.resize(aligned_face, size)
    aligned_face = aligned_face / 255.0
    aligned_face = np.transpose(aligned_face, (2, 0, 1))
    aligned_face = np.expand_dims(aligned_face, axis=0)
    aligned_face = torch.from_numpy(aligned_face).float()
    return aligned_face

def getLatent(source_face):
    return source_face.embedding


def blend_swapped_image(swapped_face, target_img, M):
    swapped_face = np.array(swapped_face) # PIL to numpy
    swapped_face = cv2.warpAffine(swapped_face, M, (target_img.shape[1], target_img.shape[0]))
    mask = np.ones_like(swapped_face) * 255
    mask = cv2.warpAffine(mask, M, (target_img.shape[1], target_img.shape[0]))

    target_img = np.array(target_img) # PIL to numpy
    swapped_face = Image.blend(Image.fromarray(target_img), Image.fromarray(swapped_face), Image.fromarray(mask).convert("L"))

    return np.array(swapped_face) # numpy to PIL


def process_images(target_image, source_image, model_path):
    args = parse_arguments()
    args.modelPath = model_path
    args.no_paste_back = False # or True, as you prefer
    args.resolution = 128
    model = load_model(args.modelPath)

    target_face_blob, M = create_target(target_image, args.resolution)
    source_latent = create_source(source_image)
    swapped_face, _ = swap_face(model, target_face_blob, source_latent)

    swapped_face = blend_swapped_image(swapped_face, target_image, M) # PIL images

    return Image.fromarray(swapped_face) # Return PIL image


with gr.Blocks() as demo:
    target_image = gr.Image(label="Target Image", type="pil") # Use PIL type
    source_image = gr.Image(label="Source Image", type="pil") # Use PIL type
    model_path = gr.Textbox(label="Model Path", value="path/to/your/model.pth") # Add model path input
    output_image = gr.Image(label="Output Image", type="pil") # Use PIL type
    btn = gr.Button("Swap Face")
    btn.click(fn=process_images, inputs=[target_image, source_image, model_path], outputs=output_image)

demo.launch() #no share = true for local running