app.py

#!/usr/bin/env python
# coding: utf-8

# # Stable Diffusion with 🤗 Diffusers




#!pip install -Uq diffusers transformers fastcore


# ## Using Stable Diffusion


import logging
from pathlib import Path
import torch
from diffusers import StableDiffusionPipeline
from PIL import Image
import numpy as np
logging.disable(logging.WARNING)


from diffusers import StableDiffusionImg2ImgPipeline
    
from fastai.vision.augment import CropPad   

import streamlit as st




imageLocation = st.empty()
with imageLocation.container():
  st.header('Animate your dream')
  st.write(f'Tap > to reveal sidebar. Select a style or artist. Enter text prompts and select the number of frames.\
  Include an optional negative prompt. Press the button to generate the animation. \
  Running on {"GPU takes 3-5 minutes." if torch.cuda.is_available() else "CPU does not work. You are advised to upgrade to (a paid) GPU after duplicating the space"}')
  st.markdown('<a style="display:inline-block" href="https://huggingface.co/spaces/sci4/AnimateYourDream?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>',unsafe_allow_html=True)
  st.image('DaliDream.gif')  


with st.sidebar:
        fn = 'Animation'   #st.text_input('Name of animation','Dali')
        style = st.text_input('Animation style (artist)','surreal Dali')
        zoom = st.checkbox('Zoom in animation',False)
        col1, col2 = st.columns(2)
        with col1:
            prompt1 = st.text_input('Prompt 1','a landscape')
            prompt2 = st.text_input('Prompt 2','weird animals')
            prompt3 = st.text_input('Prompt 3','a castle with weird animals')
        with col2:    
            frames1 = st.number_input('Frames in 1', min_value=5, max_value=10, value=5, step=1)
            frames2 = st.number_input('Frames in 2', min_value=5, max_value=10, value=5, step=1)
            frames3 = st.number_input('Frames in 3', min_value=5, max_value=10, value=5, step=1)
        negative_prompt = st.text_input('Negative prompt','text')
        

prompts = [[prompt1,frames1],
           [prompt2,frames2],
           [prompt3,frames3]]




def zoomIn(im,scale = 0.97):
    size = im.size[0]
    return im.crop_pad(int(size*scale)).resize((size,size))

def fade(im0,im1,steps=20):
    """Fade from one image to another"""
    return [Image.fromarray(((1-i/steps)*np.array(im0)+i/steps*np.array(im1)).astype(np.uint8)) for i in range(steps)]

def makeMovie(prompts, style='', negative_prompt='', scale = (512-4)/512,mix_factor=0.01,strength=0.5,guidance_scale=7,num_inference_steps=50):
  # Create an initial image then iterate
  with st.spinner('Be patient, it takes about a minute to generate the initial image on a GPU, but is likely to time out on CPU.'):    
    pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4",
                                               revision="fp16", torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32)
    if torch.cuda.is_available():
      pipe = pipe.to("cuda")
    prompt1 = f'{prompts[0][0]} in the style of {style}' if style!='' else prompts[0][0]
    im1 = pipe(prompt1).images[0]
  with st.spinner('Preparing animation pipeline takes another minute on a GPU'):
    pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
      "CompVis/stable-diffusion-v1-4",revision="fp16",torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32)
    if torch.cuda.is_available():
      pipe = pipe.to("cuda")
  im = im1
  movie = [im]
  for prompt,frames in prompts:
    prompt = f'{prompt} in the style of {style}' if style!='' else prompt
    for i in range(frames):
      im = zoomIn(im,scale)
      im2 = pipe(prompt, num_images_per_prompt=1, image=im, negative_prompt=negative_prompt,
                 strength=strength, guidance_scale=guidance_scale, 
                num_inference_steps=num_inference_steps).images[0]
      if max(max(im2.getextrema()))>0:
          #st.write(i,prompt)
          im = Image.fromarray(((1-mix_factor*i)*np.array(im2)+mix_factor*i*np.array(im1)).astype(np.uint8))
          movie += [im]
          imageLocation.image(im,caption=f'{prompt} frame {i}')
  
  n = len(movie)
  extMovie = []  
  for i in range(n):
    extMovie += fade(movie[i],movie[(i+1)%n],steps=20)

  return extMovie

def create(fn,prompts,style='',negative_prompt='',zoom=True):
    st.header('Generating initial image')
    scale = (512-16)/512 if zoom else 1
    movie = makeMovie(prompts, style, negative_prompt,scale,mix_factor=0.01,strength=0.5,guidance_scale=7,num_inference_steps=50)
    with st.spinner('Final step: stitching frames together to make animation'):
        movie[0].save(f'{fn}.gif', format='GIF', append_images=movie[1:], save_all=True, duration=50, loop=0)
    imageLocation.image(open(f'{fn}.gif','rb').read(),caption=f'{fn} displays above, as soon as it has loaded')






st.sidebar.button('Generate animation',on_click=create, args=(fn,prompts,style,negative_prompt,zoom), type='primary')