前準備
http://memo.eightban.com/stable-diffusion/stable-diffusion-diffusers1
img2img StableDiffusionControlNetImg2ImgPipeline
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel,StableDiffusionControlNetImg2ImgPipeline
#
from diffusers import UniPCMultistepScheduler
from diffusers.models import AutoencoderKL
from diffusers.utils import load_image
#import torch.utils
#from controlnet_aux import PidiNetDetector,HEDdetector, MidasDetector, MLSDdetector, OpenposeDetector, PidiNetDetector, NormalBaeDetector, LineartDetector, LineartAnimeDetector, CannyDetector, ContentShuffleDetector, ZoeDetector, MediapipeFaceDetector, SamDetector, LeresDetector
from controlnet_aux.processor import Processor
from PIL import Image, ImageOps
#import cv2
#import numpy as np
from natsort import natsorted
from huggingface_hub import HfApi
from pathlib import Path
import torch
import datetime
import os
import random
import glob
#device = "cuda"
#device = "cpu"
if device == "cpu":
torch_dtype=torch.float32
else:
torch_dtype=torch.float16
device = "cuda"
#
#init_img = Image.open("/content/drive/MyDrive/Images/sunflower.png")
#init_img = load_image( "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")
file_format = "%Y%m%d_%H%M%S"
file_list1 = glob.glob(os.path.join(load_path, "*.png"))
file_list1.extend(glob.glob(os.path.join(load_path, "*.jpg")))
file_list1 = natsorted(file_list1)
file_list2 = glob.glob(os.path.join(controlnet_path, "*.png"))
file_list2.extend(glob.glob(os.path.join(controlnet_path, "*.jpg")))
file_list2 = natsorted(file_list2)
if controlnet_image_loop==False :
file_list = file_list1
file_listx = file_list2
else:
file_listx = file_list1
file_list = file_list2
if not file_list2 :
file_list = file_list1
file_listx = file_list2
#vae = AutoencoderKL.from_pretrained(vae)
#画像生成に使うスケジューラー
#scheduler = EulerDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler")
#scheduler = DPMSolverMultistepScheduler.from_pretrained(model_id, subfolder="scheduler")
#canny = CannyDetector()
#openpose = OpenposeDetector.from_pretrained('lllyasviel/ControlNet')
#hed = HEDdetector.from_pretrained('lllyasviel/Annotators')
#hed = HEDdetector.from_pretrained('lllyasviel/ControlNet')
if controlnet_preprocessor_id != "":
processor = Processor(controlnet_preprocessor_id)
#controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-openpose", torch_dtype=torch.float16)
#controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-scribble", torch_dtype=torch.float16)
#controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
#
controlnet = ControlNetModel.from_pretrained(controlnet_processor_id, torch_dtype=torch_dtype)
#パイプラインの作成
#pipe = StableDiffusionControlNetPipeline.from_pretrained(model_id,
#scheduler=scheduler,#
if controlnet_image_loop==True and not file_list1:
pipe = StableDiffusionControlNetPipeline.from_pretrained(model_id,
controlnet=controlnet,
# vae=vae,
#custom_pipeline="lpw_stable_diffusion",
safety_checker=None,
torch_dtype=torch_dtype)
else:
pipe = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(model_id, #scheduler=scheduler,
controlnet=controlnet,
# vae=vae,
#custom_pipeline="lpw_stable_diffusion",
safety_checker=None,
torch_dtype=torch_dtype)
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
#
pipe.enable_model_cpu_offload()
#
pipe.to(device)
#NSFW規制を無効化する
if pipe.safety_checker is not None:
pipe.safety_checker = lambda images, **kwargs: (images, False)
if seed is None or seed == -1:
init_Seed = random.randint(0, 2147483647)
else:
init_Seed = seed
def resize_image(image, target_width, target_height):
original_width, original_height = image.size
original_aspect_ratio = original_width / original_height
target_aspect_ratio = target_width / target_height
aspect_ratio_difference = original_aspect_ratio / target_aspect_ratio
if aspect_ratio_difference > 1:
new_width = int(target_width * aspect_ratio_difference)
new_height = target_height
else:
new_width = target_width
new_height = int(target_height / aspect_ratio_difference)
resized_image = image.resize((new_width, new_height))
padded_image = ImageOps.pad(resized_image, (target_width, target_height), color='white')
return padded_image
idx2 = 0
for list_path in file_list:
idx2 += 1
if controlnet_image_loop==True :
if controlnet_first_image==True :
if idx2 > 1:
break
if controlnet_image_loop==False :
img_path = list_path
infile_name = os.path.basename(img_path)
initfile_path = os.path.join(controlnet_path, infile_name)
if not os.path.exists(initfile_path) :
initfile_path = img_path
if controlnet_first_image==True :
if file_listx:
initfile_path = file_listx[0]
else:
if file_listx:
img_path = file_listx[0]
else:
img_path = ""
initfile_path = list_path
if file_list1:
open_img = Image.open(img_path)
open_img = resize_image(open_img,width,height)
# open_img = Image.open(img_path).convert("RGB")
# canny_image = canny(open_img)
# openpose_image = openpose(open_img)
# scribble_image = hed(open_img, scribble=True)
init_img = Image.open(initfile_path)
if controlnet_image_loop==True and not file_list1:
pass
else:
init_img = resize_image(init_img,width,height)
if controlnet_preprocessor_id != "":
init_img = processor(init_img)
controlnet_save_path = f"/content/output/controlnet"
controlnet_image_name = os.path.basename(initfile_path)
controlnet_image_name_no_extension = os.path.splitext(controlnet_image_name)[0]
controlnet_image_name_extension = os.path.splitext(controlnet_image_name)[1]
controlnet_image_name = controlnet_image_name_no_extension + f".png"
controlnet_save_pathname = os.path.join(controlnet_save_path, controlnet_image_name)
#
init_img.save(controlnet_save_pathname)
image_name = os.path.basename(img_path)
image_name_no_extension = os.path.splitext(image_name)[0]
image_name_extension = os.path.splitext(image_name)[1]
if controlnet_image_loop==True and not file_list1:
# 現在の日本時間を取得
jst_dattetime = datetime.datetime.now(datetime.timezone(datetime.timedelta(hours=9)))
image_name_no_extension = jst_dattetime.strftime(file_format)
idx = 0
while idx < int(batch_count):
#generator
mSeed = init_Seed + idx
generator = torch.Generator(device=device).manual_seed(mSeed)
#images = []
if controlnet_image_loop==True and not file_list1:
image = pipe(prompt,
image=init_img,
negative_prompt=negative_prompt,
width=width, height=height, generator=generator,
controlnet_conditioning_scale=controlnet_conditioning_scale,
guidance_scale=CFG_scale, num_inference_steps=Steps,
guess_mode=guess_mode
).images[0]
else:
image = pipe(prompt,
image=open_img,
control_image=init_img,
negative_prompt=negative_prompt,
width=width, height=height, generator=generator,
strength=strength,
controlnet_conditioning_scale=controlnet_conditioning_scale,
guidance_scale=CFG_scale, num_inference_steps=Steps,
guess_mode=guess_mode
).images[0]
#出力する画像の名前を生成する
#outfile_name = (jst_dattetime.strftime(file_format)+ "_" + str(mSeed)+ "-" + str(idx))
outfile_name = (image_name_no_extension+ "_" + controlnet_image_name_no_extension + "_" + str(mSeed)+ "-" + str(idx))
image_name = outfile_name + f".png"
#画像を保存する
save_pathname = os.path.join(save_path, image_name)
image.save(save_pathname)
idx += 1