app.py

from diffusers import StableDiffusionPipeline
import gradio as gr
import requests
import base64
from PIL import Image, PngImagePlugin
from io import BytesIO
import torch
import os

# required for stable difussion
auth_token = os.environ.get("auth_token") #in secret space
hf_writer = gr.HuggingFaceDatasetSaver(auth_token, "flagged-find-the-next-james-bond")

pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", use_auth_token=auth_token)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
pipe = pipe.to(device)

def encode_pil_to_base64(pil_image):
  ''' From: https://github.com/gradio-app/gradio/blob/main/gradio/processing_utils.py'''
  with BytesIO() as output_bytes:

      # Copy any text-only metadata
      use_metadata = False
      metadata = PngImagePlugin.PngInfo()
      for key, value in pil_image.info.items():
          if isinstance(key, str) and isinstance(value, str):
              metadata.add_text(key, value)
              use_metadata = True

      pil_image.save(
          output_bytes, "PNG", pnginfo=(metadata if use_metadata else None)
      )
      bytes_data = output_bytes.getvalue()
  base64_str = str(base64.b64encode(bytes_data), "utf-8")
  return "data:image/png;base64," + base64_str

def decode_base64_to_image(encoding):
  ''' From: https://github.com/gradio-app/gradio/blob/main/gradio/processing_utils.py'''
  content = encoding.split(";")[1]
  image_encoded = content.split(",")[1]
  return Image.open(BytesIO(base64.b64decode(image_encoded)))

def improve_image(img, scale=2):
  ''' Improves an input image using GFP-GAN
  Inputs
  img (PIL): image to improve
  scale (int): scale factor for new image
  Output
  Improved image. If the request to GFPGAN is unsuccesful, it returns
  a black image.
  '''
  url = "https://hf.space/embed/NotFungibleIO/GFPGAN/+/api/predict"
  request_objt = {"data":[encode_pil_to_base64(img),'v1.3', scale]}
  try:
    imp_img = decode_base64_to_image(requests.post(url, json=request_objt).json()['data'][0])
  except AttributeError:
    return Image.new('RGB', size=(512, 512))
  return imp_img

def generate(celebrity, movie, guidance, improve_flag, scale): # add scale as var
  prompt = f"A movie poster of {celebrity} in {movie}."
  image = pipe(prompt, guidance=guidance).images[0]
  if improve_flag:
    image = improve_image(image, scale=scale)
  return image

movie_options = movie_options = ["Casino Royale (2006)",
                "Quantum of Solace (2008)",
                "Skyfall (2012)",
                "Spectre (2015)",
                "No Time To Die (2021)"]
title = "Find the next James Bond"
description = "Write the name of a celebrity, and pick a Bond movie from the dropdown menu.\
               This will generate a new movie poster (inspired by the chosen movie)\
               with the new celebrity in it. See below for explanation of the\
               input variables (guidance, improve and scale). Some candidates are Tom Hardy, Idris Elba, Henry Cavill, Suranne Jones, Lashana Lynch, and Robert Pattinson."
article= "Inputs explained: \n Guidance: a lower value will show more random results, but with better image quality. Improve and scale: if selected, the image resolution will be increased and faces will be higher quality with the eyes and facial features better defined), also the image size (originally 512x512) will be scaled by the chosen factor. Uses [GFP-GAN](https://github.com/TencentARC/GFPGAN)."

demo = gr.Interface(
    fn=generate,
    inputs=[gr.Textbox(value="Tom Hardy"), 
            gr.Dropdown(movie_options, value="Skyfall (2012)"), 
            gr.Slider(1, 20, value=7.5, step=0.5),
            gr.Checkbox(label="Improve and scale? (takes longer to process)"),
            gr.Slider(1, 3, value=1, step=0.5)
            ],
    outputs='image',
    title=title,
    description=description,
    article=article,
    allow_flagging="manual",
    flagging_options=["wrong celebrity", "wrong movie", "face not shown", "nonsensical output"],
    flagging_callback=hf_writer
) 

demo.launch()