Nosana: アクセスしやすいGPUでStable Diffusionモデルを実行する

%pip install diffusers transformers accelerate python-dotenv Pillow -q

# Install ipywidgets for tqdm progress bars in Jupyter
%pip install ipywidgets --upgrade -q

import torch
from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image, StableDiffusionDepth2ImgPipeline
from PIL import Image
import requests
from dotenv import load_dotenv
import os

# Create an .env file and echo your token into the file
load_dotenv()
api_key = os.getenv("HUGGINGFACE_ACCESS_TOKEN")

# Check for GPU availability
device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

print(f"Using device: {device}")

import time
from huggingface_hub import login
from diffusers import AutoPipelineForText2Image
import torch
from IPython.display import display

# Authenticate with Hugging Face Hub
login(token=api_key)

# # Load the text-to-image pipeline with authentication
# txt2img_pipe = AutoPipelineForText2Image.from_pretrained(
#     "stable-diffusion-v1-5/stable-diffusion-v1-5", 
#     torch_dtype=torch.float32,   # assuming you're using GPU
#     use_safetensors=True,
#     token=api_key
# ).to("cuda")   

# Load the text-to-image pipeline with authentication
txt2img_pipe = AutoPipelineForText2Image.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5", 
    torch_dtype=torch.float32,   # CPU must use float32
    use_safetensors=True,
    token=api_key
).to("cpu") # change to cpu or mps

# Define the prompt
prompt = "A cinematic shot of a cat in a library, wearing a monocle, detailed, 8k"
negative_prompt = "blurry, low quality, deformed, ugly"

# Measure time
start_time = time.time()

image = txt2img_pipe(
    prompt=prompt, 
    negative_prompt=negative_prompt, 
    num_inference_steps=30, 
    guidance_scale=7.5
).images[0]

end_time = time.time()
elapsed_time = end_time - start_time

print(f"Image generated in {elapsed_time:.2f} seconds")

# ✅ Display inline in Jupyter
display(image)

# Load the image-to-image pipeline
img2img_pipe = AutoPipelineForImage2Image.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5", 
    torch_dtype=torch_dtype, 
    use_safetensors=True
).to(device)

# Load an initial image from a URL
url = "https://upload.wikimedia.org/wikipedia/commons/thumb/e/ea/Van_Gogh_-_Starry_Night_-_Google_Art_Project.jpg/1280px-Van_Gogh_-_Starry_Night_-_Google_Art_Project.jpg"
init_image = Image.open(requests.get(url, stream=True).raw).resize((768, 600))

prompt = "A futuristic cityscape at night, high-tech, neon lights, in the style of Blade Runner"

# Generate the new image
image = img2img_pipe(
    prompt=prompt,
    image=init_image,
    strength=0.75,
    guidance_scale=7.5
).images[0]

# Display the transformed image
image

# Load the depth-to-image pipeline
depth_pipe = StableDiffusionDepth2ImgPipeline.from_pretrained(
    "stabilityai/stable-diffusion-2-depth",
    torch_dtype=torch_dtype,
).to(device)

# Use the same initial image
url = "https://upload.wikimedia.org/wikipedia/commons/thumb/e/ea/Van_Gogh_-_Starry_Night_-_Google_Art_Project.jpg/1280px-Van_Gogh_-_Starry_Night_-_Google_Art_Project.jpg"
init_image = Image.open(requests.get(url, stream=True).raw).resize((768, 600))

prompt = "Two majestic wolves howling at the moon, fantasy art, detailed"

# Generate the new image using depth information
image = depth_pipe(
    prompt=prompt, 
    image=init_image, 
    strength=0.7
).images[0]

# Display the result
image