app.py

from diffusers import AutoPipelineForText2Image, EulerAncestralDiscreteScheduler
from diffusers import UniPCMultistepScheduler
import gradio as gr
import torch
from PIL import Image
import random
import os
from huggingface_hub import hf_hub_download
import torch
from torch import autocast
from safetensors import safe_open
from compel import Compel, ReturnedEmbeddingsType
from safetensors.torch import load_file
import spaces

model_id = "aipicasso/emix-0-4-turbo"
auth_token=os.environ["ACCESS_TOKEN"]
#adapter_id = "latent-consistency/lcm-lora-sdxl"
#adapter_id_2 = "manual.safetensors"

scheduler = EulerAncestralDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler",token=auth_token)

pipe = AutoPipelineForText2Image.from_pretrained(
  model_id,
  torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
  scheduler=scheduler,
  token=auth_token)

#pipe = AutoPipelineForText2Image.from_pretrained(
#    model_id, 
#    torch_dtype=torch.float16,
#    use_auth_token=auth_token
#)
#pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)

pipe=pipe.to("cuda")
#pipe.load_lora_weights(adapter_id)
#pipe.load_lora_weights(adapter_id_2)
#pipe.fuse_lora()

pipe.enable_freeu(s1=0.9, s2=0.2, b1=1.3, b2=1.4)
#pipe.enable_freeu(s1=0.6, s2=0.4, b1=1.1, b2=1.2)
#pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)

state_dict = load_file("unaestheticXLv31.safetensors")
pipe.load_textual_inversion(state_dict["clip_g"], token="unaestheticXLv31", text_encoder=pipe.text_encoder_2, tokenizer=pipe.tokenizer_2)
pipe.load_textual_inversion(state_dict["clip_l"], token="unaestheticXLv31", text_encoder=pipe.text_encoder, tokenizer=pipe.tokenizer)

compel = Compel(tokenizer=[pipe.tokenizer, pipe.tokenizer_2] , 
                text_encoder=[pipe.text_encoder, pipe.text_encoder_2], 
                returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED, 
                requires_pooled=[False, True])

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

@spaces.GPU
def inference(prompt, guidance, steps, seed=0, neg_prompt="", disable_auto_prompt_correction=False):
  global pipe
    
  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)

  height=768
  width=768

  print(prompt,neg_prompt)
    
  return txt_to_img(prompt, neg_prompt, guidance, steps, width, height, generator), None

def auto_prompt_correction(prompt_ui,neg_prompt_ui,disable_auto_prompt_correction):
    # auto prompt correction
    prompt=str(prompt_ui)
    neg_prompt=str(neg_prompt_ui)
    prompt=prompt.lower()
    neg_prompt=neg_prompt.lower()
    
    if(disable_auto_prompt_correction):
        return prompt, neg_prompt

    if(prompt=="" and neg_prompt==""):
        prompt="1girl, smile, brown bob+++ hair, brown eyes, sunflowers, sky"
        neg_prompt=f"unaestheticXLv31, photo, deformed, realism, disfigured, low contrast, bad hand"
        return prompt, neg_prompt

    splited_prompt=prompt.replace(","," ").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"{prompt}"
            neg_prompt=f"unaestheticXLv31,{neg_prompt}, photo, deformed, realism, disfigured, low contrast, bad hand" 
            return prompt, neg_prompt
            
    animal_words=["cat","dog","bird","pigeon","rabbit","bunny","horse"]
    for word in animal_words:
        if( word in splited_prompt):
            prompt=f"{prompt}, 4k, detailed"
            neg_prompt=f"{neg_prompt},unaestheticXLv31" 
            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"{prompt}, 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):
    conditioning, pooled = compel([prompt, neg_prompt])

    result = pipe(
#        prompt_embeds=conditioning[0:1],
#        pooled_prompt_embeds=pooled[0:1], 
#        negative_prompt_embeds=conditioning[1:2], 
#        negative_pooled_prompt_embeds=pooled[1:2],
        prompt=prompt,
        negative=neg_prompt,
        num_inference_steps = int(steps),
        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>Emix 0.4 Turbo Demo</h1>
              </div>
              <p>
               Demo for Emix 0.4 Turbo<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 : 1girl, cool+, smile--, colorful long hair, colorful eyes, stars, night, pastel color, transparent+
              </p>
              <p>
              sample prompt2 : 1boy, focus, wavy short hair, blue eyes, black shirt, white background, simple background
              </p>
              <p>
              <a style="display:inline-block" href="https://huggingface.co/spaces/aipicasso/emi-latest-demo?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a> to say goodbye from waiting for the generating.
              </p>
            </div>
        """
    )
    with gr.Row():
        
        with gr.Column(scale=55):
          with gr.Group():
              with gr.Row():
                prompt = gr.Textbox(label="Prompt", show_label=False, max_lines=2,placeholder="[your prompt]")
                generate = gr.Button(value="Generate")

              image_out = gr.Image()
          error_output = gr.Markdown()

        with gr.Column(scale=45):
            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.")
                
              with gr.Row():
                guidance = gr.Slider(label="Guidance scale", value=1.5, maximum=10, step=0.1)
                steps = gr.Slider(label="Steps", value=8, minimum=1, maximum=20, step=1)
    
              seed = gr.Slider(0, 2147483647, label='Seed (0 = random)', value=0, step=1)

    inputs = [prompt, guidance, steps, seed, neg_prompt, disable_auto_prompt_correction]

    outputs = [image_out, error_output]
    prompt.submit(inference, inputs=inputs, outputs=outputs)
    generate.click(inference, inputs=inputs, outputs=outputs)
    
demo.queue()
demo.launch()