from flask import Flask, request, jsonify, stream_with_context, send_file, send_from_directory, Response import asyncio import torch import shutil import os import sys from time import strftime from src.utils.preprocess import CropAndExtract from src.test_audio2coeff import Audio2Coeff from src.facerender.animate import AnimateFromCoeff from src.generate_batch import get_data from src.generate_facerender_batch import get_facerender_data # from src.utils.init_path import init_path import tempfile from openai import OpenAI import elevenlabs from elevenlabs import set_api_key, generate, play, clone, Voice, VoiceSettings import uuid import time from PIL import Image import moviepy.editor as mp import requests import json import pickle # from dotenv import load_dotenv from concurrent.futures import ProcessPoolExecutor, as_completed, ThreadPoolExecutor from stream_server import add_video,HLS_DIR, generate_m3u8 import math # Load environment variables from .env file # load_dotenv() # Initialize ProcessPoolExecutor for parallel processing executor = ThreadPoolExecutor(max_workers=2) torch.cuda.empty_cache() class AnimationConfig: def __init__(self, driven_audio_path, source_image_path, result_folder,pose_style,expression_scale,enhancer,still,preprocess,ref_pose_video_path, image_hardcoded): self.driven_audio = driven_audio_path self.source_image = source_image_path self.ref_eyeblink = None self.ref_pose = None self.checkpoint_dir = './checkpoints' self.result_dir = result_folder self.pose_style = pose_style self.batch_size = 2 self.expression_scale = expression_scale self.input_yaw = None self.input_pitch = None self.input_roll = None self.enhancer = enhancer self.background_enhancer = None self.cpu = False self.face3dvis = False self.still = still self.preprocess = preprocess self.verbose = False self.old_version = False self.net_recon = 'resnet50' self.init_path = None self.use_last_fc = False self.bfm_folder = './checkpoints/BFM_Fitting/' self.bfm_model = 'BFM_model_front.mat' self.focal = 1015. self.center = 112. self.camera_d = 10. self.z_near = 5. self.z_far = 15. self.device = 'cuda' self.image_hardcoded = image_hardcoded app = Flask(__name__) from flask_cors import CORS CORS(app,origins=["*"]) TEMP_DIR = None start_time = None audio_chunks = [] preprocessed_data = None args = None unique_id = None m3u8_path = None audio_duration = None driven_audio_path = None app.config['temp_response'] = None app.config['generation_thread'] = None app.config['text_prompt'] = None app.config['final_video_path'] = None app.config['final_video_duration'] = None # Global paths dir_path = os.path.dirname(os.path.realpath(__file__)) current_root_path = dir_path path_of_lm_croper = os.path.join(current_root_path, 'checkpoints', 'shape_predictor_68_face_landmarks.dat') path_of_net_recon_model = os.path.join(current_root_path, 'checkpoints', 'epoch_20.pth') dir_of_BFM_fitting = os.path.join(current_root_path, 'checkpoints', 'BFM_Fitting') wav2lip_checkpoint = os.path.join(current_root_path, 'checkpoints', 'wav2lip.pth') audio2pose_checkpoint = os.path.join(current_root_path, 'checkpoints', 'auido2pose_00140-model.pth') audio2pose_yaml_path = os.path.join(current_root_path, 'src', 'config', 'auido2pose.yaml') audio2exp_checkpoint = os.path.join(current_root_path, 'checkpoints', 'auido2exp_00300-model.pth') audio2exp_yaml_path = os.path.join(current_root_path, 'src', 'config', 'auido2exp.yaml') free_view_checkpoint = os.path.join(current_root_path, 'checkpoints', 'facevid2vid_00189-model.pth.tar') # Function for running the actual task (using preprocessed data) def process_chunk(audio_chunk, preprocessed_data, args): print("Entered Process Chunk Function") global audio2pose_checkpoint, audio2pose_yaml_path, audio2exp_checkpoint, audio2exp_yaml_path, wav2lip_checkpoint global free_view_checkpoint if args.preprocess == 'full': mapping_checkpoint = os.path.join(current_root_path, 'checkpoints', 'mapping_00109-model.pth.tar') facerender_yaml_path = os.path.join(current_root_path, 'src', 'config', 'facerender_still.yaml') else: mapping_checkpoint = os.path.join(current_root_path, 'checkpoints', 'mapping_00229-model.pth.tar') facerender_yaml_path = os.path.join(current_root_path, 'src', 'config', 'facerender.yaml') first_coeff_path = preprocessed_data["first_coeff_path"] crop_pic_path = preprocessed_data["crop_pic_path"] crop_info_path = "/home/user/app/preprocess_data/crop_info.json" with open(crop_info_path , "rb") as f: crop_info = json.load(f) print(f"Loaded existing preprocessed data") print("first_coeff_path",first_coeff_path) print("crop_pic_path",crop_pic_path) print("crop_info",crop_info) torch.cuda.empty_cache() batch = get_data(first_coeff_path, audio_chunk, args.device, ref_eyeblink_coeff_path=None, still=args.still) audio_to_coeff = Audio2Coeff(audio2pose_checkpoint, audio2pose_yaml_path, audio2exp_checkpoint, audio2exp_yaml_path, wav2lip_checkpoint, args.device) coeff_path = audio_to_coeff.generate(batch, args.result_dir, args.pose_style, ref_pose_coeff_path=None) # Further processing with animate_from_coeff using the coeff_path animate_from_coeff = AnimateFromCoeff(free_view_checkpoint, mapping_checkpoint, facerender_yaml_path, args.device) torch.cuda.empty_cache() data = get_facerender_data(coeff_path, crop_pic_path, first_coeff_path, audio_chunk, args.batch_size, args.input_yaw, args.input_pitch, args.input_roll, expression_scale=args.expression_scale, still_mode=args.still, preprocess=args.preprocess) torch.cuda.empty_cache() print("Will Enter Animation") result, base64_video, temp_file_path, _ = animate_from_coeff.generate(data, args.result_dir, args.source_image, crop_info, enhancer=args.enhancer, background_enhancer=args.background_enhancer, preprocess=args.preprocess) # video_clip = mp.VideoFileClip(temp_file_path) # duration = video_clip.duration app.config['temp_response'] = base64_video app.config['final_video_path'] = temp_file_path # app.config['final_video_duration'] = duration torch.cuda.empty_cache() return base64_video, temp_file_path def create_temp_dir(): return tempfile.TemporaryDirectory() def save_uploaded_file(file, filename,TEMP_DIR): print("Entered save_uploaded_file") unique_filename = str(uuid.uuid4()) + "_" + filename file_path = os.path.join(TEMP_DIR.name, unique_filename) file.save(file_path) return file_path def custom_cleanup(temp_dir): # Iterate over the files and directories in TEMP_DIR for filename in os.listdir(temp_dir): file_path = os.path.join(temp_dir, filename) if os.path.isdir(file_path): shutil.rmtree(file_path) else: os.remove(file_path) print(f"Deleted: {file_path}") torch.cuda.empty_cache() import gc gc.collect() # def get_audio_duration(audio_path): # audio_clip = mp.AudioFileClip(audio_path) # duration_in_seconds = audio_clip.duration # audio_clip.close() # Don't forget to close the clip # return duration_in_seconds def generate_audio(voice_cloning, voice_gender, text_prompt): print("generate_audio") if voice_cloning == 'no': if voice_gender == 'male': voice = 'echo' print('Entering Audio creation using elevenlabs') set_api_key('92e149985ea2732b4359c74346c3daee') audio = generate(text = text_prompt, voice = "Daniel", model = "eleven_multilingual_v2",stream=True, latency=4) with tempfile.NamedTemporaryFile(suffix=".mp3", prefix="text_to_speech_",dir=TEMP_DIR.name, delete=False) as temp_file: for chunk in audio: temp_file.write(chunk) driven_audio_path = temp_file.name print('driven_audio_path',driven_audio_path) print('Audio file saved using elevenlabs') else: voice = 'nova' print('Entering Audio creation using whisper') response = client.audio.speech.create(model="tts-1-hd", voice=voice, input = text_prompt) print('Audio created using whisper') with tempfile.NamedTemporaryFile(suffix=".wav", prefix="text_to_speech_",dir=TEMP_DIR.name, delete=False) as temp_file: driven_audio_path = temp_file.name response.write_to_file(driven_audio_path) print('Audio file saved using whisper') elif voice_cloning == 'yes': set_api_key('92e149985ea2732b4359c74346c3daee') # voice = clone(name = "User Cloned Voice", # files = [user_voice_path] ) voice = Voice(voice_id="CEii8R8RxmB0zhAiloZg",name="Marc",settings=VoiceSettings( stability=0.71, similarity_boost=0.5, style=0.0, use_speaker_boost=True),) audio = generate(text = text_prompt, voice = voice, model = "eleven_multilingual_v2",stream=True, latency=4) with tempfile.NamedTemporaryFile(suffix=".mp3", prefix="cloned_audio_",dir=TEMP_DIR.name, delete=False) as temp_file: for chunk in audio: temp_file.write(chunk) driven_audio_path = temp_file.name print('driven_audio_path',driven_audio_path) # audio_duration = get_audio_duration(driven_audio_path) # print('Total Audio Duration in seconds',audio_duration) return driven_audio_path def run_preprocessing(args): global path_of_lm_croper, path_of_net_recon_model, dir_of_BFM_fitting first_frame_dir = os.path.join(args.result_dir, 'first_frame_dir') os.makedirs(first_frame_dir, exist_ok=True) fixed_temp_dir = "/home/user/app/preprocess_data/" os.makedirs(fixed_temp_dir, exist_ok=True) preprocessed_data_path = os.path.join(fixed_temp_dir, "preprocessed_data.pkl") if os.path.exists(preprocessed_data_path) and args.image_hardcoded == "yes": print("Loading preprocessed data...") with open(preprocessed_data_path, "rb") as f: preprocessed_data = pickle.load(f) print("Loaded existing preprocessed data from:", preprocessed_data_path) return preprocessed_data client = OpenAI(api_key="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") def openai_chat_avatar(text_prompt): response = client.chat.completions.create( model="gpt-4o-mini", messages=[{"role": "system", "content": "Ensure answers are concise, human-like, and clear while maintaining quality. Use the fewest possible words, avoiding unnecessary articles, prepositions, and adjectives. Responses should be short but still address the question thoroughly without being verbose.Keep them to one sentence only"}, {"role": "user", "content": f"Hi! I need help with something. {text_prompt}"}, ], max_tokens = len(text_prompt) + 300 # Use the length of the input text # temperature=0.3, # stop=["Translate:", "Text:"] ) return response def split_audio(audio_path, TEMP_DIR, chunk_duration): audio_clip = mp.AudioFileClip(audio_path) total_duration = audio_clip.duration print("split_audio duration:",total_duration) number_of_chunks = math.ceil(total_duration / chunk_duration) print("Number of audio chunks:",number_of_chunks) audio_chunks = [] for i in range(number_of_chunks): start_time = i * chunk_duration end_time = min(start_time + chunk_duration, total_duration) chunk = audio_clip.subclip(start_time, end_time) # Create a temporary file for the chunk with tempfile.NamedTemporaryFile(suffix=f"_chunk_{start_time}-{end_time}.wav", prefix="audio_chunk_", dir=TEMP_DIR.name, delete=False) as temp_file: chunk_path = temp_file.name chunk.write_audiofile(chunk_path) # Specify codec if needed audio_chunks.append((start_time, chunk_path)) audio_clip.close() # Close the audio clip to release resources return audio_chunks, total_duration # def extract_order_from_path(temp_file_path): # match = re.search(r'videostream(\d+)', temp_file_path) # return int(match.group(1)) if match else -1 # Return -1 if no match is found, handle appropriately. # Generator function to yield chunk results as they are processed def generate_chunks(audio_chunks, preprocessed_data, args, m3u8_path, audio_duration, start_time): global TEMP_DIR future_to_chunk = {executor.submit(process_chunk, chunk[1], preprocessed_data, args): chunk[0] for chunk in audio_chunks} processed_chunks = {chunk[0]: None for chunk in audio_chunks} print("processed_chunks:",processed_chunks) yielded_count = 1 try: for chunk_idx, future in enumerate(as_completed(future_to_chunk)): idx = future_to_chunk[future] try: base64_video, temp_file_path = future.result() processed_chunks[idx] = temp_file_path for expected_start_time in sorted(processed_chunks.keys()): if processed_chunks[expected_start_time] is not None: add_video(processed_chunks[expected_start_time], m3u8_path, audio_duration) end_time = time.time() elapsed_time = end_time - start_time event_data = json.dumps({ 'start_time': expected_start_time, 'video_index': yielded_count, 'elapsed_time': elapsed_time }) yield f"data: {event_data}\n\n" processed_chunks[expected_start_time] = None yielded_count += 1 else: break except Exception as e: yield f"Task for chunk {idx} failed: {e}\n" finally: if TEMP_DIR: #close_m3u8(m3u8_path) custom_cleanup(TEMP_DIR.name) def close_m3u8(m3u8_path: str): try: with open(m3u8_path, 'a') as m3u8_file: m3u8_file.write('#EXT-X-ENDLIST\n') print(f"Closed m3u8 file with end tag: {m3u8_path}") except Exception as e: print(f"Error closing m3u8 file: {e}") @app.route("/run", methods=['POST']) def parallel_processing(): global start_time, driven_audio_path global audio_chunks, preprocessed_data, args, m3u8_path, audio_duration start_time = time.time() global TEMP_DIR TEMP_DIR = create_temp_dir() global unique_id unique_id = str(uuid.uuid4()) print('request:',request.method) try: if request.method == 'POST': # source_image = request.files['source_image'] image_path = '/home/user/app/images/marc_smile_enhanced.jpg' source_image = Image.open(image_path) text_prompt = request.form['text_prompt'] print('Input text prompt: ',text_prompt) text_prompt = text_prompt.strip() if not text_prompt: return jsonify({'error': 'Input text prompt cannot be blank'}), 400 voice_cloning = request.form.get('voice_cloning', 'yes') image_hardcoded = request.form.get('image_hardcoded', 'no') chat_model_used = request.form.get('chat_model_used', 'openai') target_language = request.form.get('target_language', 'original_text') print('target_language',target_language) pose_style = int(request.form.get('pose_style', 1)) expression_scale = float(request.form.get('expression_scale', 1)) enhancer = request.form.get('enhancer', None) voice_gender = request.form.get('voice_gender', 'male') still_str = request.form.get('still', 'False') still = still_str.lower() == 'false' print('still', still) preprocess = request.form.get('preprocess', 'crop') print('preprocess selected: ',preprocess) # ref_pose_video = request.files.get('ref_pose', None) response = openai_chat_avatar(text_prompt) text_prompt = response.choices[0].message.content.strip() app.config['text_prompt'] = text_prompt print('Final output text prompt using openai: ',text_prompt) source_image_path = save_uploaded_file(source_image, 'source_image.png',TEMP_DIR) print(source_image_path) driven_audio_path = generate_audio(voice_cloning, voice_gender, text_prompt) save_dir = tempfile.mkdtemp(dir=TEMP_DIR.name) result_folder = os.path.join(save_dir, "results") os.makedirs(result_folder, exist_ok=True) ref_pose_video_path = None args = AnimationConfig(driven_audio_path=driven_audio_path, source_image_path=source_image_path, result_folder=result_folder, pose_style=pose_style, expression_scale=expression_scale,enhancer=enhancer,still=still,preprocess=preprocess,ref_pose_video_path=ref_pose_video_path, image_hardcoded=image_hardcoded) preprocessed_data = run_preprocessing(args) # chunk_duration = 3 # print(f"Splitting the audio into {chunk_duration}-second chunks...") # audio_chunks, audio_duration = split_audio(driven_audio_path, TEMP_DIR, chunk_duration=chunk_duration) # print(f"Audio has been split into {len(audio_chunks)} chunks: {audio_chunks}") start_time = 0 audio_clip = mp.AudioFileClip(driven_audio_path) audio_duration = audio_clip.duration audio_chunks.append((start_time, driven_audio_path)) os.makedirs('lives', exist_ok=True) print("Entering generate m3u8") m3u8_path = f'lives/{unique_id}.m3u8' #generate_m3u8(audio_duration, m3u8_path) return jsonify({'video_url': f'{unique_id}.m3u8'}), 200 except Exception as e: app.logger.error(f"An error occurred: {e}") return jsonify({'status': 'error', 'message': str(e)}), 500 @app.route("/stream", methods=["GET"]) def stream_results(): global audio_chunks, preprocessed_data, args, m3u8_path, audio_duration, start_time print("audio_chunks",audio_chunks) print("preprocessed_data",preprocessed_data) print("args",args) try: return Response(stream_with_context(generate_chunks(audio_chunks, preprocessed_data, args, m3u8_path, audio_duration, start_time)),content_type='text/event-stream') except Exception as e: return jsonify({'status': 'error', 'message': str(e)}), 500 @app.route("/live_stream/", methods=['GET']) async def get_concatenated_playlist(playlist: str): """ Endpoint to serve the concatenated HLS playlist. Returns: FileResponse: The concatenated playlist file. """ if playlist.endswith('.ts'): playlist_path = os.path.join('hls_videos', playlist) else: playlist_path = os.path.join('lives', playlist) if not os.path.exists(playlist_path): return jsonify({'status': 'error', "msg":"Playlist not found"}), 404 return send_file(playlist_path, mimetype='application/vnd.apple.mpegurl') # @app.route("/live_stream/", methods=["GET"]) # def live_stream(filename): # return send_from_directory(directory="hls_videos", filename=filename) @app.route("/health", methods=["GET"]) def health_status(): response = {"online": "true"} return jsonify(response) if __name__ == '__main__': app.run(debug=True)