Spaces:

LuChengTHU
/

dpmsolver_sdm

Runtime error

App Files Files Community

dpmsolver_sdm / app.py

LuChengTHU

change requirements

39d2dd2 over 1 year ago

raw history blame

No virus

11.2 kB

	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,
	predict_epsilon=True,
	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)