import gradio as gr import moviepy.video.io.ImageSequenceClip from PIL import Image from pydub import AudioSegment # Import everything needed to edit video clips from moviepy.editor import * import numpy as np import mutagen from mutagen.mp3 import MP3 import cv2 def resize(img_list): print("** inside resize **") print('Entity-Images generated by multimodal interface are:',img_list) resize_img_list = [] for item in img_list: im = Image.open(item) imResize = im.resize((256,256), Image.ANTIALIAS) resize_img_list.append(np.array(imResize)) print('Type of elements in the image list:',type(resize_img_list[0])) return resize_img_list def merge_audio_video(entities_num, resize_img_list, text_input): print("** inside merge aud vid **") print('Type of image list variable: ',type(resize_img_list)) print('Type of elements in the image list: ',type(resize_img_list[0])) #Convert text to speech using facebook's latest model from HF hub speech = text2speech(text_input) print('Back in merge_audio_video') print('Type of speech variable : ',type(speech)) print('Type of Audio file: ',speech) wav_audio = AudioSegment.from_file(speech, "flac") #("/content/gdrive/My Drive/AI/audio1.flac", "flac") #convert flac to mp3 audio format print('COnverting flac format to mp3 using AudioSegment object:', type(wav_audio)) wav_audio.export("audio.mp3", format="mp3") #("/content/gdrive/My Drive/AI/audio1.mp3", format="mp3") print('flac audio converted to mp3 audio' ) print('now getting duration of this mp3 audio' ) #getting audio clip's duration audio_length = int(MP3("audio.mp3").info.length) print('Audio length is :',audio_length) #Calculate the desired frame per second based on given audio length and entities identified fps= entities_num / audio_length #length of audio file fps = float(format(fps, '.5f')) print('Based on number of entities/images and audio length, FPS is set as : ',fps) #String a list of images into a video and write to memory clip = moviepy.video.io.ImageSequenceClip.ImageSequenceClip(resize_img_list, fps=fps) clip.write_videofile('my_vid_tmp.mp4') print('video clip created successfully from images') # loading video file print('Starting video and audio merge') videoclip = VideoFileClip('my_vid_tmp.mp4') #("/content/gdrive/My Drive/AI/my_video1.mp4") print('loading video-clip') # loading audio file audioclip = AudioFileClip('audio.mp3') #.subclip(0, 15) print('loading mp3-format audio') # adding audio to the video clip mergedclip = videoclip.set_audio(audioclip) print('video and audio merged successfully') #Getting size and frame count of merged video file print('Getting size and frame count of merged video file') duration = mergedclip.duration frame_count = mergedclip.fps print('duration is:',duration) print('frame count :', frame_count) return mergedclip fastspeech = gr.Interface.load("huggingface/facebook/fastspeech2-en-ljspeech") def text2speech(text): print('** inside testtospeech **') print('Loading the model through :',type(fastspeech)) print(fastspeech) speech = fastspeech(text) print('Type of variable in which file is stored:',type(speech)) print('Type of Audio file generated :',speech) return speech def engine(text_input): print(" ** Inside Enngine **") #Extract entities from text ner = gr.Interface.load("huggingface/flair/ner-english-ontonotes-large") entities = ner(text_input) entities = [tupl for tupl in entities if None not in tupl] entities_num = len(entities) #Generate images using multimodelart's space for each entity identified above img_list = [] for ent in entities: img = gr.Interface.load("spaces/multimodalart/latentdiffusion")(ent[0],'50','256','256','1',10)[0] img_list.append(img) print('img_list size:',len(img_list)) #Resizing all images produced to same size resize_img_list = resize(img_list) print('back from resize into engine') #Merge video and audio created above mergedclip = merge_audio_video(entities_num, resize_img_list, text_input) print('\n Back in engine') print(' Merged clip type :',type(mergedclip)) print('Writing the merged video clip to a video file') mergedclip.to_videofile('mergedvideo.mp4') print('mergedvideo.mp4 created') print('################################ Single Run Completed ##############################') return 'mergedvideo.mp4' app = gr.Interface(engine, gr.inputs.Textbox(lines=5, label="Input Text"), gr.outputs.Video(type=None, label='Final Merged video'), description="
Firstly, the Demo generates speech from input-text using facebook's fastspeech2-en-ljspeech from HF hub.
Then, takes the input-text and extracts the entities in it using Flair NER model from HF Hub.
Then, generate images using Multimodalart Space for every entity separately.
Creates a video by stringing all the entity-images together.
Lastly, Fuses the AI generated audio and video together to create a coherent movie for you to watch.

A fun little app that lets you turn your text to video (well, in some ways atleast :) ). More the entities in your text, More time to build the output, More fun to watch.
Please expect build time of around 10-20 seconds per entity. For instance, in the third and largest example there are 13 entities as per the NER model used here.
" , examples=["On April 17th Sunday George celebrated Easter. He is staying at Empire State building with his parents.", "George is a citizen of Canada and speaks English and French fluently. His role model is the former president Obama. " , "On April 17th Sunday George celebrated Easter. He is staying at Empire State building with his parents. He is a citizen of Canada and speaks English and French fluently. His role model is former president Obama. He got 1000 dollar from his mother to visit Disney World and to buy new iPhone mobile. George likes watching Game of Thrones.", "April is the month of Easter weekend. Visit places like Statue of Liberty with friends. Take at least 200 dollars in cash with you. Use Android phone to find places in Newyork City."], title="Generate Audio & Video from Text", article="
For best results, make sure to enter a text that has entities listed on model card for flair/ner-english-ontonotes-large. Some examples of type of entities that will be helpful are - Date values, event names, building names, languages, locations, money value, organization names, famous people names, products and so on.
Also note that, this Space loads the most awesome Multimodalart space as a gradio interface, hence if the latter space is down former goes down too.

Who owns the videos produced by this demo?

(Borrowing this from multimodalart spaces) Definetly not me! Probably you do. I say probably because the Copyright discussion about AI generated art is ongoing. So it may be the case that everything produced here falls automatically into the public domain. But in any case it is either yours or is in the public domain.
" ).launch(enable_queue=True, debug=True)