run.py

import os
import gradio as gr
import cv2
import numpy as np
from insightface.app import FaceAnalysis
import tempfile
from moviepy.editor import VideoFileClip

# ===== Configuration précoce pour éviter les erreurs =====

# Désactiver les warnings d'Albumentations (optionnel)
os.environ["NO_ALBUMENTATIONS_UPDATE"] = "1"

# Rediriger les chemins problématiques
os.environ['MPLCONFIGDIR'] = '/tmp/matplotlib'
os.environ['FONTCONFIG_PATH'] = '/tmp/fontconfig'
os.environ['FONTCONFIG_FILE'] = '/tmp/fontconfig/fonts.conf'

# Créer les répertoires temporaires nécessaires
os.makedirs(os.environ['MPLCONFIGDIR'], exist_ok=True)
os.makedirs(os.environ['FONTCONFIG_PATH'], exist_ok=True)

# Créer un fichier minimal de configuration Fontconfig si nécessaire
with open(os.environ['FONTCONFIG_FILE'], 'w') as f:
    f.write('''<?xml version="1.0"?>
<!DOCTYPE fontconfig SYSTEM "fonts.dtd">
<fontconfig>
</fontconfig>''')

# Forcer InsightFace à utiliser un répertoire accessible
os.environ['INSIGHTFACE_ROOT'] = '/tmp/.insightface'
os.environ["ORT_DISABLE_CUDA"] = "1"  # Désactiver CUDA si GPU indisponible

# ===== Fonction de swap de visage =====

def swap_face(source_face, target_face, frame):
    src_emb = source_face.normed_embedding
    tgt_bbox = target_face.bbox.astype(int)
    resized_face = cv2.resize(source_face.img, (tgt_bbox[2] - tgt_bbox[0], tgt_bbox[3] - tgt_bbox[1]))

    mask = np.zeros_like(resized_face)
    center = (mask.shape[1] // 2, mask.shape[0] // 2)
    radius = int(min(mask.shape) * 0.45)
    cv2.circle(mask, center, radius, (255, 255, 255), -1)
    mask = cv2.GaussianBlur(mask, (15, 15), 5)

    center = ((tgt_bbox[0] + tgt_bbox[2]) // 2, (tgt_bbox[1] + tgt_bbox[3]) // 2)
    result = cv2.seamlessClone(resized_face, frame, mask, center, cv2.NORMAL_CLONE)
    return result


def process_video(source_img, target_video):
    try:
        face_app = FaceAnalysis(name="buffalo_l", root="/tmp/.insightface")
        face_app.prepare(ctx_id=0, det_size=(640, 640))

        source_faces = face_app.get(source_img)
        if not source_faces:
            raise ValueError("Aucun visage trouvé dans l'image source.")
        source_face = source_faces[0]

        temp_output = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)

        cap = cv2.VideoCapture(target_video)
        fps = cap.get(cv2.CAP_PROP_FPS)
        frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
        frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

        fourcc = cv2.VideoWriter_fourcc(*'avc1')  # Codec compatible H.264
        out = cv2.VideoWriter(temp_output.name, fourcc, fps, (frame_width, frame_height))

        while True:
            ret, frame = cap.read()
            if not ret:
                break

            target_faces = face_app.get(frame)
            for face in target_faces:
                frame = swap_face(source_face, face, frame)

            out.write(frame)

        cap.release()
        out.release()

        print(f"Taille de la vidéo temporaire : {os.path.getsize(temp_output.name)} octets")

        # Réencodage final pour compatibilité Gradio
        clip = VideoFileClip(temp_output.name)
        final_path = tempfile.mktemp(suffix=".mp4")
        clip.write_videofile(final_path, codec="libx264", audio_codec="aac", verbose=False, logger=None)

        return final_path

    except Exception as e:
        print(f"Erreur lors du traitement : {str(e)}")
        return None


# ===== Interface Gradio =====

demo = gr.Interface(
    fn=process_video,
    inputs=[
        gr.Image(label="Visage Source", type="numpy"),
        gr.Video(label="Vidéo Cible"),
    ],
    outputs=gr.Video(label="Vidéo Résultat"),
    title="🎬 FaceSwap Pro",
    description="Échangez des visages dans une vidéo.",
    allow_flagging="never"
)


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