app.py

from diffusers import AutoencoderKL, UNet2DConditionModel, StableDiffusionPipeline, StableDiffusionImg2ImgPipeline, DPMSolverMultistepScheduler
import gradio as gr
import torch
from PIL import Image
import os

scheduler = DPMSolverMultistepScheduler(
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    num_train_timesteps=1000,
    trained_betas=None,
    prediction_type="epsilon",
    thresholding=False,
    algorithm_type="dpmsolver++",
    solver_type="midpoint",
    lower_order_final=True,
)

class Model:
    def __init__(self, name, path, prefix):
        self.name = name
        self.path = path
        self.prefix = prefix
        self.pipe_t2i = None
        self.pipe_i2i = None

models = [
     Model("Stable-Diffusion-v1.4", "CompVis/stable-diffusion-v1-4", "The 1.4 version of official stable-diffusion"),
     Model("Waifu", "hakurei/waifu-diffusion", "anime style"),
]

last_mode = "txt2img"
current_model = models[0]
current_model_path = current_model.path

auth_token = os.getenv("HUGGING_FACE_HUB_TOKEN")

print(f"Is CUDA available: {torch.cuda.is_available()}")

if torch.cuda.is_available():
  vae = AutoencoderKL.from_pretrained(current_model.path, subfolder="vae", torch_dtype=torch.float16, use_auth_token=auth_token)
  for model in models:
    try:
        unet = UNet2DConditionModel.from_pretrained(model.path, subfolder="unet", torch_dtype=torch.float16, use_auth_token=auth_token)
        model.pipe_t2i = StableDiffusionPipeline.from_pretrained(model.path, unet=unet, vae=vae, torch_dtype=torch.float16, scheduler=scheduler, use_auth_token=auth_token)
        model.pipe_i2i = StableDiffusionImg2ImgPipeline.from_pretrained(model.path, unet=unet, vae=vae, torch_dtype=torch.float16, scheduler=scheduler, use_auth_token=auth_token)
    except:
        models.remove(model)
  pipe = models[0].pipe_t2i
  pipe = pipe.to("cuda")

else:
  vae = AutoencoderKL.from_pretrained(current_model.path, subfolder="vae", use_auth_token=auth_token)
  for model in models:
    try:
        unet = UNet2DConditionModel.from_pretrained(model.path, subfolder="unet", use_auth_token=auth_token)
        model.pipe_t2i = StableDiffusionPipeline.from_pretrained(model.path, unet=unet, vae=vae, scheduler=scheduler, use_auth_token=auth_token)
        model.pipe_i2i = StableDiffusionImg2ImgPipeline.from_pretrained(model.path, unet=unet, vae=vae, scheduler=scheduler, use_auth_token=auth_token)
    except:
        models.remove(model)
  pipe = models[0].pipe_t2i
  pipe = pipe.to("cpu")

device = "GPU 🔥" if torch.cuda.is_available() else "CPU 🥶"

def inference(model_name, prompt, guidance, steps, width=512, height=512, seed=0, img=None, strength=0.5, neg_prompt=""):

  global current_model
  for model in models:
    if model.name == model_name:
      current_model = model
      model_path = current_model.path

  generator = torch.Generator('cuda' if torch.cuda.is_available() else 'cpu').manual_seed(seed) if seed != 0 else None

  if img is not None:
    return img_to_img(model_path, prompt, neg_prompt, img, strength, guidance, steps, width, height, generator)
  else:
    return txt_to_img(model_path, prompt, neg_prompt, guidance, steps, width, height, generator)

def txt_to_img(model_path, prompt, neg_prompt, guidance, steps, width, height, generator=None):

    global last_mode
    global pipe
    global current_model_path
    if model_path != current_model_path or last_mode != "txt2img":
        current_model_path = model_path

        pipe.to("cpu")
        pipe = current_model.pipe_t2i

        if torch.cuda.is_available():
          pipe = pipe.to("cuda")
        last_mode = "txt2img"

    prompt = current_model.prefix + prompt
    result = pipe(
      prompt,
      negative_prompt = neg_prompt,
      # num_images_per_prompt=n_images,
      num_inference_steps = int(steps),
      guidance_scale = guidance,
      width = width,
      height = height,
      generator = generator)
    
    return replace_nsfw_images(result)

def img_to_img(model_path, prompt, neg_prompt, img, strength, guidance, steps, width, height, generator=None):

    global last_mode
    global pipe
    global current_model_path
    if model_path != current_model_path or last_mode != "img2img":
        current_model_path = model_path

        pipe.to("cpu")
        pipe = current_model.pipe_i2i
        
        if torch.cuda.is_available():
          pipe = pipe.to("cuda")
        last_mode = "img2img"

    prompt = current_model.prefix + prompt
    ratio = min(height / img.height, width / img.width)
    img = img.resize((int(img.width * ratio), int(img.height * ratio)), Image.LANCZOS)
    result = pipe(
        prompt,
        negative_prompt = neg_prompt,
        # num_images_per_prompt=n_images,
        init_image = img,
        num_inference_steps = int(steps),
        strength = strength,
        guidance_scale = guidance,
        #width = width,
        #height = height,
        generator = generator)
        
    return replace_nsfw_images(result)

def replace_nsfw_images(results):
    for i in range(len(results.images)):
      if results.nsfw_content_detected[i]:
        results.images[i] = Image.open("nsfw.png")
    return results.images[0]

css = """
  <style>
  .finetuned-diffusion-div {
      text-align: center;
      max-width: 700px;
      margin: 0 auto;
      font-family: 'IBM Plex Sans', sans-serif;
    }
    .finetuned-diffusion-div div {
      display: inline-flex;
      align-items: center;
      gap: 0.8rem;
      font-size: 1.75rem;
    }
    .finetuned-diffusion-div div h1 {
      font-weight: 900;
      margin-top: 15px;
      margin-bottom: 15px;
      text-align: center;
      line-height: 150%;
    }
    .finetuned-diffusion-div p {
      margin-bottom: 10px;
      font-size: 94%;
    }
    .finetuned-diffusion-div p a {
      text-decoration: underline;
    }
    .tabs {
      margin-top: 0px;
      margin-bottom: 0px;
    }
    #gallery {
      min-height: 20rem;
    }
    .container {
      max-width: 1000px;
      margin: auto;
      padding-top: 1.5rem;
    }
  </style>
"""
with gr.Blocks(css=css) as demo:
    gr.HTML(
        f"""
            <div class="finetuned-diffusion-div">
              <div>
                <h1>Stable-Diffusion with DPM-Solver (fastest sampler for diffusion models) </h1>
              </div>
              <br>
              <p>
              ❤️ Acknowledgement: Hardware resources of this demo are supported by HuggingFace 🤗 . Many thanks for the help!
              </p>
              <br>
              <p>
               This is a demo of sampling by DPM-Solver with two variants of Stable Diffusion models, including <a href="https://huggingface.co/CompVis/stable-diffusion-v1-4">Stable-Diffusion-v1.4</a> and <a href="https://huggingface.co/hakurei/waifu-diffusion">Waifu</a>.
              </p>
              <br>
              <p>
               <a href="https://github.com/LuChengTHU/dpm-solver">DPM-Solver</a> (Neurips 2022 Oral) is a fast high-order solver customized for diffusion ODEs, which can generate high-quality samples by diffusion models within only 10-25 steps. DPM-Solver has an analytical formulation and is very easy to use for all types of Gaussian diffusion models, and includes <a href="https://arxiv.org/abs/2010.02502">DDIM</a> as a first-order special case.
              </p>
              <p>
              We use <a href="https://github.com/huggingface/diffusers">Diffusers</a>  🧨  to implement this demo, which currently supports the multistep DPM-Solver scheduler. For more details of DPM-Solver with Diffusers, check <a href="https://github.com/huggingface/diffusers/pull/1132">this pull request</a>.
              </p>
              <br>
              <p>
              Currently, the default sampler of stable-diffusion is <a href="https://arxiv.org/abs/2202.09778">PNDM</a>, which needs 50 steps to generate high-quality samples. However, DPM-Solver can generate high-quality samples within only <span style="font-weight: bold;">20-25</span> steps, and for some samples even within <span style="font-weight: bold;">10-15</span> steps.
              </p>
              <br>
              <p>
               Running on <b>{device}</b>
              </p>
            </div>
        """
    )

    with gr.Row():
        
        with gr.Column(scale=55):
          with gr.Group():
              model_name = gr.Dropdown(label="Model", choices=[m.name for m in models], value=current_model.name)
              with gr.Row():
                prompt = gr.Textbox(label="Prompt", show_label=False, max_lines=2,placeholder="Enter prompt. Style applied automatically").style(container=False)
                generate = gr.Button(value="Generate").style(rounded=(False, True, True, False))


              image_out = gr.Image(height=512)
              # gallery = gr.Gallery(
              #     label="Generated images", show_label=False, elem_id="gallery"
              # ).style(grid=[1], height="auto")

        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")

              # n_images = gr.Slider(label="Images", value=1, minimum=1, maximum=4, step=1)

              with gr.Row():
                guidance = gr.Slider(label="Guidance scale", value=7.5, maximum=15)
                steps = gr.Slider(label="Steps", value=25, minimum=2, maximum=100, step=1)

              with gr.Row():
                width = gr.Slider(label="Width", value=512, minimum=64, maximum=1024, step=8)
                height = gr.Slider(label="Height", value=512, minimum=64, maximum=1024, step=8)

              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)

    # model_name.change(lambda x: gr.update(visible = x == models[0].name), inputs=model_name, outputs=custom_model_group)

    inputs = [model_name, prompt, guidance, steps, width, height, seed, image, strength, neg_prompt]
    prompt.submit(inference, inputs=inputs, outputs=image_out)

    generate.click(inference, inputs=inputs, outputs=image_out)


    gr.Markdown('''
      Stable-diffusion Models by [CompVis](https://huggingface.co/CompVis) and [stabilityai](https://huggingface.co/stabilityai), Waifu-diffusion models by [@hakurei](https://huggingface.co/hakurei). Most of the code of this demo are copied from [@anzorq's fintuned-diffusion](https://huggingface.co/spaces/anzorq/finetuned_diffusion/tree/main) ❤️<br>
      Space by [Cheng Lu](https://github.com/LuChengTHU). [![Twitter Follow](https://img.shields.io/twitter/follow/ChengLu05671218?label=%40ChengLu&style=social)](https://twitter.com/ChengLu05671218)
        
      ![visitors](https://visitor-badge.glitch.me/badge?page_id=LuChengTHU.dpmsolver_sdm)
    ''')

demo.queue(concurrency_count=1)
demo.launch(debug=False, share=False)