app.py

from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline, EulerAncestralDiscreteScheduler
from transformers import CLIPFeatureExtractor
import gradio as gr
import torch
from PIL import Image
import random
import os
from huggingface_hub import hf_hub_download


model_id = 'aipicasso/picasso-diffusion-1-0-demo'
auth_token=os.environ.get("ACCESS_TOKEN")

scheduler = EulerAncestralDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler", use_auth_token=auth_token)
feature_extractor = CLIPFeatureExtractor.from_pretrained(model_id, use_auth_token=auth_token)

pipe_merged = StableDiffusionPipeline.from_pretrained(
  model_id,
  torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
  scheduler=scheduler, use_auth_token=auth_token)

pipe_i2i_merged = StableDiffusionImg2ImgPipeline.from_pretrained(
  model_id,
  torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
  scheduler=scheduler,
  requires_safety_checker=False,
  safety_checker=None,
  feature_extractor=feature_extractor, use_auth_token=auth_token
)

pipe=pipe_merged.to("cuda")
pipe_i2i=pipe_i2i_merged.to("cuda")
pipe.enable_xformers_memory_efficient_attention()
pipe_i2i.enable_xformers_memory_efficient_attention()

embeddings_path=hf_hub_download(repo_id=model_id, filename="nfixer.pt", use_auth_token=auth_token)
embeddings_dict=torch.load(embeddings_path)
print(embeddings_dict)
if "string_to_param" in embeddings_dict:
    embeddings = next(iter(embeddings_dict['string_to_param'].values()))
    nfixer = ""
    for i, emb in enumerate(embeddings):
        token = f"sksd{chr(i+98)}"
        nfixer += token
        pipe.tokenizer.add_tokens(token)
        pipe.text_encoder.resize_token_embeddings(len(pipe.tokenizer))
        token_id = pipe.tokenizer.convert_tokens_to_ids(token)
        pipe.text_encoder.get_input_embeddings().weight.data[token_id] = emb
else:
    nfixer = list(embeddings_dict.keys())[0]
    embeddings = embeddings_dict[nfixer].to(pipe.text_encoder.get_input_embeddings().weight.dtype)
    pipe.tokenizer.add_tokens(placeholder_token)
    pipe.text_encoder.resize_token_embeddings(len(pipe.tokenizer))
    placeholder_token_id = pipe.tokenizer.convert_tokens_to_ids(placeholder_token)
    pipe.text_encoder.get_input_embeddings().weight.data[placeholder_token_id] = embeddings

embeddings_path=hf_hub_download(repo_id=model_id, filename="embellish2.pt", use_auth_token=auth_token)
embeddings_dict=torch.load(embeddings_path)
print(embeddings_dict)
if "string_to_param" in embeddings_dict:
    embeddings = next(iter(embeddings_dict['string_to_param'].values()))
    embellish2 = ""
    for i, emb in enumerate(embeddings):
        token = f"kskd{chr(i+98)}"
        embellish2 += token
        pipe.tokenizer.add_tokens(token)
        pipe.text_encoder.resize_token_embeddings(len(pipe.tokenizer))
        token_id = pipe.tokenizer.convert_tokens_to_ids(token)
        pipe.text_encoder.get_input_embeddings().weight.data[token_id] = emb
else:
    embellish2 = list(embeddings_dict.keys())[0]
    embeddings = embeddings_dict[embellish2].to(pipe.text_encoder.get_input_embeddings().weight.dtype)
    pipe.tokenizer.add_tokens(placeholder_token)
    pipe.text_encoder.resize_token_embeddings(len(pipe.tokenizer))
    placeholder_token_id = pipe.tokenizer.convert_tokens_to_ids(placeholder_token)
    pipe.text_encoder.get_input_embeddings().weight.data[placeholder_token_id] = embeddings

def error_str(error, title="Error"):
    return f"""#### {title}
            {error}"""  if error else ""

def inference(prompt, guidance, steps, image_size="Square", seed=0, img=None, strength=0.5, neg_prompt="", disable_auto_prompt_correction=False, image_style="Realistic", original_model=False):
  global pipe,pipe_i2i
  
  generator = torch.Generator('cuda').manual_seed(seed) if seed != 0 else None

  prompt,neg_prompt=auto_prompt_correction(prompt,neg_prompt,disable_auto_prompt_correction,image_style)

  if(image_size=="Portrait"):
      height=1024
      width=768
  elif(image_size=="Landscape"):
      height=768
      width=1024  
  elif(image_size=="Highreso."):
      height=1024
      width=1024  
  else:
      height=768
      width=768

  print(prompt,neg_prompt)
    
  try:
    if img is not None:
      return img_to_img(prompt, neg_prompt, img, strength, guidance, steps, width, height, generator), None
    else:
      return txt_to_img(prompt, neg_prompt, guidance, steps, width, height, generator), None
  except Exception as e:
    return None, error_str(e)
def auto_prompt_correction(prompt_ui,neg_prompt_ui,disable_auto_prompt_correction,image_style):
    # auto prompt correction
    prompt=str(prompt_ui)
    neg_prompt=str(neg_prompt_ui)
    prompt=prompt.lower()
    neg_prompt=neg_prompt.lower()

    if(image_style=="Animetic"):
        style="anime"
    else:
        style=f"anime,{embellish2}"
    
    if(disable_auto_prompt_correction):
        prompt=f"{style}, {prompt}"
        return prompt, neg_prompt

    if(prompt=="" and neg_prompt==""):
        prompt=f"{style}, masterpiece, portrait, a girl with flowers, good pupil, 4k, detailed"
        neg_prompt=f"{nfixer},(((deformed))), blurry, ((((bad anatomy)))), bad pupil, disfigured, poorly drawn face, mutation, mutated, (extra limb), (ugly), (poorly drawn hands), bad hands, fused fingers, messy drawing, broken legs censor, low quality, (mutated hands and fingers:1.5), (long body :1.3), (mutation, poorly drawn :1.2), ((bad eyes)), ui, error, missing fingers, fused fingers, one hand with more than 5 fingers, one hand with less than 5 fingers, one hand with more than 5 digit, one hand with less than 5 digit, extra digit, fewer digits, fused digit, missing digit, bad digit, liquid digit, long body, uncoordinated body, unnatural body, lowres, jpeg artifacts, 3d, cg, text"
        return prompt, neg_prompt

    splited_prompt=prompt.replace(","," ").replace("_"," ").split(" ")
    
    human_words=["1girl","girl","maid","maids","female","1woman","woman","girls","2girls","3girls","4girls","5girls","a couple of girls","women","1boy","boy","boys","a couple of boys","2boys","male","1man","1handsome","1bishounen","man","men","guy","guys"]
    for word in human_words:
        if( word in splited_prompt):
            prompt=f"{style}, masterpiece, {prompt}, good pupil, 4k, detailed"
            neg_prompt=f"{nfixer},(((deformed))), blurry, ((((bad anatomy)))), {neg_prompt}, bad pupil, disfigured, poorly drawn face, mutation, mutated, (extra limb), (ugly), (poorly drawn hands), bad hands, fused fingers, messy drawing, broken legs censor, low quality, (mutated hands and fingers:1.5), (long body :1.3), (mutation, poorly drawn :1.2), ((bad eyes)), ui, error, missing fingers, fused fingers, one hand with more than 5 fingers, one hand with less than 5 fingers, one hand with more than 5 digit, one hand with less than 5 digit, extra digit, fewer digits, fused digit, missing digit, bad digit, liquid digit, long body, uncoordinated body, unnatural body, lowres, jpeg artifacts, 3d, cg, text"
            return prompt, neg_prompt
            
    animal_words=["cat","dog","bird"]
    for word in animal_words:
        if( word in splited_prompt):
            prompt=f"{style}, a {prompt}, 4k, detailed"
            neg_prompt=f"{nfixer}, girl, (((deformed))), blurry, ((((bad anatomy)))), {neg_prompt}, bad pupil, disfigured, poorly drawn face, mutation, mutated, (extra limb), (ugly), (poorly drawn hands), bad hands, fused fingers, messy drawing, broken legs censor, low quality, (mutated hands and fingers:1.5), (long body :1.3), (mutation, poorly drawn :1.2), ((bad eyes)), ui, error, missing fingers, fused fingers, one hand with more than 5 fingers, one hand with less than 5 fingers, one hand with more than 5 digit, one hand with less than 5 digit, extra digit, fewer digits, fused digit, missing digit, bad digit, liquid digit, long body, uncoordinated body, unnatural body, lowres, jpeg artifacts, 3d, cg, text"
            return prompt, neg_prompt
            
    background_words=["mount fuji","mt. fuji","building", "buildings", "tokyo", "kyoto", "nara", "shibuya", "shinjuku"]
    for word in background_words:
        if( word in splited_prompt):
            prompt=f"{style}, shinkai makoto, {word}, 4k, 8k, highly detailed"
            neg_prompt=f"girl, (((deformed))), {neg_prompt}, girl, boy, photo, people, low quality, ui, error, lowres, jpeg artifacts, 2d, 3d, cg, text"
            return prompt, neg_prompt
            
    return prompt,neg_prompt
    
def txt_to_img(prompt, neg_prompt, guidance, steps, width, height, generator):
    result = pipe(
      prompt,
      negative_prompt = neg_prompt,
      num_inference_steps = int(steps),
      guidance_scale = guidance,
      width = width,
      height = height,
      generator = generator)
    
    return result.images[0]

def img_to_img(prompt, neg_prompt, img, strength, guidance, steps, width, height, generator):
    ratio = min(height / img.height, width / img.width)
    img = img.resize((int(img.width * ratio), int(img.height * ratio)), Image.LANCZOS)
    result = pipe_i2i(
        prompt,
        negative_prompt = neg_prompt,
        init_image = img,
        num_inference_steps = int(steps),
        strength = strength,
        guidance_scale = guidance,
        #width = width,
        #height = height,
        generator = generator)
        
    return result.images[0]

css = """.main-div div{display:inline-flex;align-items:center;gap:.8rem;font-size:1.75rem}.main-div div h1{font-weight:900;margin-bottom:7px}.main-div p{margin-bottom:10px;font-size:94%}a{text-decoration:underline}.tabs{margin-top:0;margin-bottom:0}#gallery{min-height:20rem}
"""
with gr.Blocks(css=css) as demo:
    gr.HTML(
        f"""
            <div class="main-div">
              <div>
                <h1>Picasso Diffusion 1.0 +α Demo</h1>
              </div>
              <p>
               Demo for <a href="https://huggingface.co/aipicasso/picasso-diffusion-1-0">Picasso Diffusion 1.0 (Comming soon)</a> + <a href="https://twitter.com/cac0e/status/1622381533892952069?s=20&t=0zTDN_4D14LP7w2eVvLzOg">cacoe's model</a>.<br>
              </p>
              <p>
              サンプル： そのままGenerateボタンを押してください。<br>
              sample : Click "Generate" button without any prompts.
              </p>
              <p>
              sample prompt1 : girl, kimono
              </p>
              <p>
              sample prompt2 : boy, armor
              </p>
              Running on {"<b>GPU 🔥</b>" if torch.cuda.is_available() else f"<b>CPU 🥶</b>. For faster inference it is recommended to <b>upgrade to GPU in <a href='https://huggingface.co/spaces/akhaliq/cool-japan-diffusion-2-1-0/settings'>Settings</a></b>"} <br>
              <!-- <a style="display:inline-block" href="https://huggingface.co/spaces/aipicasso/picasso-diffusion-latest-demo?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a> to say goodbye from waiting for the generating. -->
            </div>
        """
    )
    with gr.Row():
        
        with gr.Column(scale=55):
          with gr.Group():
              with gr.Row():
                image_style=gr.Radio(["Realistic","Animetic"])
                image_style.show_label=False
                image_style.value="Realistic"
                prompt = gr.Textbox(label="Prompt", show_label=False, max_lines=2,placeholder="[your prompt]").style(container=False)
                generate = gr.Button(value="Generate").style(rounded=(False, True, True, False))

              image_out = gr.Image(height=768,width=768)
          error_output = gr.Markdown()

        with gr.Column(scale=45):
          with gr.Tab("Options"):
            with gr.Group():
              neg_prompt = gr.Textbox(label="Negative prompt", placeholder="What to exclude from the image")
              disable_auto_prompt_correction = gr.Checkbox(label="Disable auto prompt corretion.")
              #original_model = gr.Checkbox(label="Change the model into the original model.")
              with gr.Row():
                image_size=gr.Radio(["Portrait","Landscape","Square","Highreso."])
                image_size.show_label=False
                image_size.value="Square"
                
              with gr.Row():
                guidance = gr.Slider(label="Guidance scale", value=7.5, maximum=15)
                steps = gr.Slider(label="Steps", value=20, minimum=2, maximum=75, step=1)

              seed = gr.Slider(0, 2147483647, label='Seed (0 = random)', value=0, step=1)

          with gr.Tab("Image to image"):
              with gr.Group():
                image = gr.Image(label="Image", height=256, tool="editor", type="pil")
                strength = gr.Slider(label="Transformation strength", minimum=0, maximum=1, step=0.01, value=0.5)
                  
    inputs = [prompt, guidance, steps, image_size, seed, image, strength, neg_prompt, disable_auto_prompt_correction,image_style]#, original_model]

    outputs = [image_out, error_output]
    prompt.submit(inference, inputs=inputs, outputs=outputs)
    generate.click(inference, inputs=inputs, outputs=outputs,api_name="generate")

demo.queue(concurrency_count=1)
demo.launch()