JoyType / functions.py
jiangchen16
initial commit
3c3804b
raw
history blame
18.2 kB
import json
import os
import cv2
import random
import numpy as np
import gradio as gr
import torch
from zhipuai import ZhipuAI
from pytorch_lightning import seed_everything
from pprint import pprint
from PIL import Image, ImageDraw, ImageFont
from diffusers import (
ControlNetModel,
StableDiffusionControlNetPipeline,
)
from diffusers import (
DDIMScheduler,
PNDMScheduler,
EulerAncestralDiscreteScheduler,
DPMSolverMultistepScheduler,
EulerDiscreteScheduler,
LMSDiscreteScheduler,
HeunDiscreteScheduler
)
from controlnet_aux import (
PidiNetDetector,
HEDdetector
)
BBOX_MAX_NUM = 8
BBOX_INI_NUM = 0
MAX_LENGTH = 20
device = 'cuda'
pipeline = None
pre_pipeline = None
model_root = os.getenv('REPO_ROOT')
scheduler_root = f'{model_root}/Scheduler'
model_list =[
'JoyType.v1.0', 'RevAnimated-animation-动漫', 'GhostMix-animation-动漫',
'rpg.v5-fantasy_realism-奇幻写实', 'midjourneyPapercut-origami-折纸版画',
'dvarchExterior-architecture-建筑', 'awpainting.v13-portrait-人物肖像'
]
chn_example_dict = {
'漂亮的风景照,很多山峰,清澈的湖水': 'beautiful landscape, many peaks, clear lake',
'画有玫瑰的卡片,明亮的背景': 'a card with roses, bright background',
'一张关于健康教育的卡片,上面有一些文字,有一些食物图标,背景里有一些水果喝饮料的图标,且背景是模糊的': \
'a card for health education, with some writings on it, '
'food icons on the card, some fruits and drinking in the background, blur background '
}
match_dict = {
'JoyType.v1.0': 'JoyType-v1-1M',
'RevAnimated-animation-动漫': 'rev-animated-v1-2-2',
'GhostMix-animation-动漫': 'GhostMix_V2.0',
'rpg.v5-fantasy_realism-奇幻写实': 'rpg_v5',
'midjourneyPapercut-origami-折纸版画': 'midjourneyPapercut_v1',
'dvarchExterior-architecture-建筑': 'dvarchExterior',
'awpainting.v13-portrait-人物肖像': 'awpainting_v13'
}
font_list = [
'CHN-华文行楷',
'CHN-华文新魏',
'CHN-清松手写体',
'CHN-巴蜀墨迹',
'CHN-雷盖体',
'CHN-演示夏行楷',
'CHN-鸿雷板书简体',
'CHN-斑马字类',
'CHN-青柳隶书',
'CHN-辰宇落雁体',
'CHN-宅家麦克笔',
'ENG-Playwrite',
'ENG-Okesip',
'ENG-Shrikhand',
'ENG-Nextstep',
'ENG-Filthyrich',
'ENG-BebasNeue',
'ENG-Gloock',
'ENG-Lemon',
'RUS-Automatons',
'RUS-MKyrill',
'RUS-Alice',
'RUS-Caveat',
'KOR-ChosunGs',
'KOR-Dongle',
'KOR-GodoMaum',
'KOR-UnDotum',
'JPN-GlTsukiji',
'JPN-Aoyagireisyosimo',
'JPN-KouzanMouhitu',
'JPN-Otomanopee'
]
def change_settings(base_model):
if base_model == model_list[0]:
return gr.update(value=20), gr.update(value=7.5), gr.update(value='PNDM')
elif base_model == model_list[1]:
return gr.update(value=30), gr.update(value=8.5), gr.update(value='Euler')
elif base_model == model_list[2]:
return gr.update(value=32), gr.update(value=8.5), gr.update(value='Euler')
elif base_model == model_list[3]:
return gr.update(value=20), gr.update(value=7.5), gr.update(value='DPM')
elif base_model == model_list[4]:
return gr.update(value=25), gr.update(value=6.5), gr.update(value='Euler')
elif base_model == model_list[5]:
return gr.update(value=25), gr.update(value=8.5), gr.update(value='Euler')
elif base_model == model_list[6]:
return gr.update(value=25), gr.update(value=7), gr.update(value='DPM')
else:
pass
def update_box_num(choice):
update_list_1 = [] # checkbox
update_list_2 = [] # font
update_list_3 = [] # text
update_list_4 = [] # bounding box
for i in range(BBOX_MAX_NUM):
if i < choice:
update_list_1.append(gr.update(value=True))
update_list_2.append(gr.update(visible=True))
update_list_3.append(gr.update(visible=True))
update_list_4.extend([gr.update(visible=False) for _ in range(4)])
else:
update_list_1.append(gr.update(value=False))
update_list_2.append(gr.update(visible=False, value='CHN-华文行楷'))
update_list_3.append(gr.update(visible=False, value=''))
update_list_4.extend([
gr.update(visible=False, value=0.4),
gr.update(visible=False, value=0.4),
gr.update(visible=False, value=0.2),
gr.update(visible=False, value=0.2)
])
return *update_list_1, *update_list_2, *update_list_3, *update_list_4
def load_box_list(example_id, choice):
with open(f'templates/{example_id}.json', 'r') as f:
info = json.load(f)
update_list1 = []
update_list2 = []
update_list3 = []
update_list4 = []
for i in range(BBOX_MAX_NUM):
visible = info['visible'][i]
pos = info['pos'][i * 4: (i + 1) * 4]
update_list1.append(gr.update(value=visible))
update_list2.append(gr.update(value=info['font'][i], visible=visible))
update_list3.append(gr.update(value=info['text'][i], visible=visible))
update_list4.extend([
gr.update(value=pos[0]),
gr.update(value=pos[1]),
gr.update(value=pos[2]),
gr.update(value=pos[3])
])
return *update_list1, *update_list2, \
*update_list3, *update_list4, gr.update(value=-1)
def re_edit():
global BBOX_MAX_NUM
update_list = []
for i in range(BBOX_MAX_NUM):
update_list.extend([gr.update(value=0.4), gr.update(value=0.4), gr.update(value=0.2),
gr.update(value=0.2)])
return *update_list, \
gr.Image(
value=create_canvas(),
label='Rect Position', elem_id='MD-bbox-rect-t2i',
show_label=False, visible=True
), \
gr.Slider(value=512), gr.Slider(value=512)
def resize_w(w, img):
return cv2.resize(img, (w, img.shape[0]))
def resize_h(h, img):
return cv2.resize(img, (img.shape[1], h))
def create_canvas(w=512, h=512, c=3, line=5):
image = np.full((h, w, c), 200, dtype=np.uint8)
for i in range(h):
if i % (w // line) == 0:
image[i, :, :] = 150
for j in range(w):
if j % (w // line) == 0:
image[:, j, :] = 150
image[h // 2 - 8:h // 2 + 8, w // 2 - 8:w // 2 + 8, :] = [200, 0, 0]
return image
def canny(img):
low_threshold = 64
high_threshold = 100
img = cv2.Canny(img, low_threshold, high_threshold)
img = img[:, :, None]
img = np.concatenate([img, img, img], axis=2)
return Image.fromarray(img)
def judge_overlap(coord_list1, coord_list2):
judge = coord_list1[0] < coord_list2[2] and coord_list1[2] > coord_list2[0] \
and coord_list1[1] < coord_list2[3] and coord_list1[3] > coord_list2[1]
return judge
def parse_render_list(box_list, shape, box_num):
width = shape[0]
height = shape[1]
polygons = []
font_names = []
texts = []
valid_list = box_list[:box_num]
pos_list = box_list[box_num: 5 * box_num]
font_name_list = box_list[5 * box_num: 6 * box_num]
text_list = box_list[6 * box_num: 7 * box_num]
empty_flag = False
print(font_name_list, text_list)
for i, valid in enumerate(valid_list):
if valid:
pos = pos_list[i * 4: (i + 1) * 4]
top_left_x = int(pos[0] * width)
top_left_y = int(pos[1] * height)
w = int(pos[2] * width)
h = int(pos[3] * height)
font_name = str(font_name_list[i])
text = str(text_list[i])
if text == '':
empty_flag = True
text = 'JoyType'
if w <= 0 or h <= 0:
gr.Warning(f'Area of the box{i + 1} cannot be zero!')
return [], False
polygon = [
top_left_x,
top_left_y,
w, h
]
try:
assert font_name in font_list
font_name = font_name.split('-')[-1]
except Exception as e:
gr.Warning('Please choose a correct font!')
return [], False
polygons.append(polygon)
font_names.append(font_name.split('-')[-1])
texts.append(text)
if empty_flag:
gr.Warning('Null strings will be filled automatically!')
for i in range(len(polygons)):
for j in range(i + 1, len(polygons)):
if judge_overlap(
[polygons[i][0], polygons[i][1], polygons[i][0] + polygons[i][2], polygons[i][1] + polygons[i][3]],
[polygons[j][0], polygons[j][1], polygons[j][0] + polygons[j][2], polygons[j][1] + polygons[j][3]]
):
gr.Warning('Find overlapping boxes!')
return [], False
render_list = []
for i in range(len(polygons)):
text_dict = {}
text_dict['text'] = texts[i]
text_dict['polygon'] = polygons[i]
text_dict['font_name'] = font_names[i]
render_list.append(text_dict)
return render_list, True
def render_all_text(render_list, shape, threshold=512):
width = shape[0]
height = shape[1]
board = Image.new('RGB', (width, height), 'black')
for text_dict in render_list:
text = text_dict['text']
polygon = text_dict['polygon']
font_name = text_dict['font_name']
if len(text) > MAX_LENGTH:
text = text[:MAX_LENGTH]
gr.Warning(f'{text}... exceeds the maximum length {MAX_LENGTH} and has been cropped.')
w, h = polygon[2:]
vert = True if w < h else False
image4ratio = Image.new('RGB', (1024, 1024), 'black')
draw = ImageDraw.Draw(image4ratio)
try:
font = ImageFont.truetype(f'./font/{font_name}.ttf', encoding='utf-8', size=50)
except FileNotFoundError:
font = ImageFont.truetype(f'./font/{font_name}.otf', encoding='utf-8', size=50)
if not vert:
draw.text(xy=(0, 0), text=text, font=font, fill='white')
_, _, _tw, _th = draw.textbbox(xy=(0, 0), text=text, font=font)
_th += 1
else:
_tw, y_c = 0, 0
for c in text:
draw.text(xy=(0, y_c), text=c, font=font, fill='white')
_l, _t, _r, _b = font.getbbox(c)
_tw = max(_tw, _r - _l)
y_c += _b
_th = y_c + 1
ratio = (_th * w) / (_tw * h)
text_img = image4ratio.crop((0, 0, _tw, _th))
x_offset, y_offset = 0, 0
if 0.8 <= ratio <= 1.2:
text_img = text_img.resize((w, h))
elif ratio < 0.75:
resize_h = int(_th * (w / _tw))
text_img = text_img.resize((w, resize_h))
y_offset = (h - resize_h) // 2
else:
resize_w = int(_tw * (h / _th))
text_img = text_img.resize((resize_w, h))
x_offset = (w - resize_w) // 2
board.paste(text_img, (polygon[0] + x_offset, polygon[1] + y_offset))
return board
def load_pipeline(model_name, scheduler_name):
controlnet_path = os.path.join(model_root, f'{match_dict["JoyType.v1.0"]}')
model_path = os.path.join(model_root, model_name)
scheduler_name = scheduler_name.lower()
if scheduler_name == 'pndm':
scheduler = PNDMScheduler.from_pretrained(scheduler_root, subfolder='pndm')
if scheduler_name == 'lms':
scheduler = LMSDiscreteScheduler.from_pretrained(scheduler_root, subfolder='lms')
if scheduler_name == 'euler':
scheduler = EulerDiscreteScheduler.from_pretrained(scheduler_root, subfolder='euler')
if scheduler_name == 'dpm':
scheduler = DPMSolverMultistepScheduler.from_pretrained(scheduler_root, subfolder='dpm')
if scheduler_name == 'ddim':
scheduler = DDIMScheduler.from_pretrained(scheduler_root, subfolder='ddim')
if scheduler_name == 'heun':
scheduler = HeunDiscreteScheduler.from_pretrained(scheduler_root, subfolder='heun')
if scheduler_name == 'euler-ancestral':
scheduler = EulerAncestralDiscreteScheduler.from_pretrained(scheduler_root, subfolder='euler-ancestral')
controlnet = ControlNetModel.from_pretrained(
controlnet_path,
subfolder='controlnet',
torch_dtype=torch.float32
)
pipeline = StableDiffusionControlNetPipeline.from_pretrained(
model_path,
scheduler=scheduler,
controlnet=controlnet,
torch_dtype=torch.float32,
).to(device)
return pipeline
def preprocess_prompt(prompt):
client = ZhipuAI(api_key=os.getenv('ZHIPU_API_KEY'))
response = client.chat.completions.create(
model="glm-4-0520",
messages=[
{
'role': 'system',
'content': '''
Stable Diffusion是一款利用深度学习的文生图模型,支持通过使用提示词来产生新的图像,描述要包含或省略的元素。
我在这里引入Stable Diffusion算法中的Prompt概念,又被称为提示符。这里的Prompt通常可以用来描述图像,
他由普通常见的单词构成,最好是可以在数据集来源站点找到的著名标签(比如Ddanbooru)。
下面我将说明Prompt的生出步骤,这里的Prompt主要用于描述人物。在Prompt的生成中,你需要通过提示词来描述 人物属性,主题,外表,情绪,衣服,姿势,视角,动作,背景。
用英语单词或短语甚至自然语言的标签来描述,并不局限于我给你的单词。然后将你想要的相似的提示词组合在一起,请使用英文半角,做分隔符,每个提示词不要带引号,并将这些按从最重要到最不重要的顺序 排列。
另外请您注意,永远在每个 Prompt的前面加上引号里的内容,
“(((best quality))),(((ultra detailed))),(((masterpiece))),illustration,” 这是高质量的标志。
人物属性中,1girl表示你生成了一个女孩,2girls表示生成了两个女孩,一次。另外再注意,Prompt中不能带有-和_。
可以有空格和自然语言,但不要太多,单词不能重复。只返回Prompt。
'''
},
{
'role': 'user',
'content': prompt
}
],
temperature=0.5,
max_tokens=2048,
top_p=1,
stream=False,
)
if response:
glm = []
glm_return_list = response.choices
for item in glm_return_list:
glm.append(item.message.content)
return {'flag': 1, 'data': glm}
else:
return {'flag': 0, 'data': {}}
def process(
num_samples,
a_prompt,
n_prompt,
conditioning_scale,
cfg_scale,
inference_steps,
seed,
usr_prompt,
rect_img,
base_model,
scheduler_name,
box_num,
*box_list
):
if usr_prompt == '':
gr.Warning('Must input a prompt!')
return None, gr.Markdown('error')
if seed == -1:
seed = random.randint(0, 2147483647)
seed_everything(seed)
# Support Chinese Input
if usr_prompt in chn_example_dict.keys():
usr_prompt = chn_example_dict[usr_prompt]
else:
for ch in usr_prompt:
if '\u4e00' <= ch <= '\u9fff':
data = preprocess_prompt(usr_prompt)
if data['flag'] == 1:
usr_prompt = data['data'][0][1: -1]
else:
gr.Warning('Something went wrong while translating your prompt, please try again.')
return None, gr.Markdown('error')
break
shape = (rect_img.shape[1], rect_img.shape[0])
render_list, flag = parse_render_list(box_list, shape, box_num)
if flag:
render_img = render_all_text(render_list, shape)
else:
return None, gr.Markdown('error')
model_name = match_dict[base_model]
render_img = canny(np.array(render_img))
w, h = render_img.size
global pipeline, pre_pipeline
if pre_pipeline != model_name or pipeline is None:
pre_pipeline = model_name
pipeline = load_pipeline(model_name, scheduler_name)
batch_render_img = [render_img for _ in range(num_samples)]
batch_prompt = [f'{usr_prompt}, {a_prompt}' for _ in range(num_samples)]
batch_n_prompt = [n_prompt for _ in range(num_samples)]
images = pipeline(
batch_prompt,
negative_prompt=batch_n_prompt,
image=batch_render_img,
controlnet_conditioning_scale=float(conditioning_scale),
guidance_scale=float(cfg_scale),
width=w,
height=h,
num_inference_steps=int(inference_steps),
).images
return images, gr.Markdown(f'{seed}, {usr_prompt}, {box_list}')
def draw_example(box_list, color, id):
board = Image.fromarray(create_canvas())
w, h = board.size
draw = ImageDraw.Draw(board, mode='RGBA')
visible = box_list[:BBOX_MAX_NUM]
pos = box_list[BBOX_MAX_NUM: 5 * BBOX_MAX_NUM]
font = box_list[5 * BBOX_MAX_NUM: 6 * BBOX_MAX_NUM]
text = box_list[6 * BBOX_MAX_NUM:]
info = {
'visible': list(visible),
'pos': list(pos),
'font': list(font),
'text': list(text)
}
with open(f'templates/{id}.json', 'w') as f:
json.dump(info, f)
for i in range(BBOX_MAX_NUM):
if visible[i] is True:
polygon = pos[i * 4: (i + 1) * 4]
print(polygon)
left = w * polygon[0]
top = h * polygon[1]
right = left + w * polygon[2]
bottom = top + h * polygon[3]
draw.rectangle([left, top, right, bottom], outline=color[i][0], fill=color[i][1], width=3)
board.save(f'./examples/{id}.png')
if __name__ == '__main__':
pass