kolcontrl / app.py
fantos's picture
Update app.py
78d6af0 verified
import spaces
import random
import torch
import cv2
import gradio as gr
import numpy as np
from huggingface_hub import snapshot_download
from transformers import CLIPVisionModelWithProjection, CLIPImageProcessor
from diffusers.utils import load_image
from kolors.pipelines.pipeline_controlnet_xl_kolors_img2img import StableDiffusionXLControlNetImg2ImgPipeline
from kolors.models.modeling_chatglm import ChatGLMModel
from kolors.models.tokenization_chatglm import ChatGLMTokenizer
from kolors.models.controlnet import ControlNetModel
from diffusers import AutoencoderKL
from kolors.models.unet_2d_condition import UNet2DConditionModel
from diffusers import EulerDiscreteScheduler
from PIL import Image
from annotator.midas import MidasDetector
from annotator.dwpose import DWposeDetector
from annotator.util import resize_image, HWC3
from transformers import pipeline
device = "cuda"
ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors")
ckpt_dir_depth = snapshot_download(repo_id="Kwai-Kolors/Kolors-ControlNet-Depth")
ckpt_dir_canny = snapshot_download(repo_id="Kwai-Kolors/Kolors-ControlNet-Canny")
ckpt_dir_pose = snapshot_download(repo_id="Kwai-Kolors/Kolors-ControlNet-Pose")
text_encoder = ChatGLMModel.from_pretrained(f'{ckpt_dir}/text_encoder', torch_dtype=torch.float16).half().to(device)
tokenizer = ChatGLMTokenizer.from_pretrained(f'{ckpt_dir}/text_encoder')
vae = AutoencoderKL.from_pretrained(f"{ckpt_dir}/vae", revision=None).half().to(device)
scheduler = EulerDiscreteScheduler.from_pretrained(f"{ckpt_dir}/scheduler")
unet = UNet2DConditionModel.from_pretrained(f"{ckpt_dir}/unet", revision=None).half().to(device)
controlnet_depth = ControlNetModel.from_pretrained(f"{ckpt_dir_depth}", revision=None).half().to(device)
controlnet_canny = ControlNetModel.from_pretrained(f"{ckpt_dir_canny}", revision=None).half().to(device)
controlnet_pose = ControlNetModel.from_pretrained(f"{ckpt_dir_pose}", revision=None).half().to(device)
pipe_depth = StableDiffusionXLControlNetImg2ImgPipeline(
vae=vae,
controlnet=controlnet_depth,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
force_zeros_for_empty_prompt=False
)
pipe_canny = StableDiffusionXLControlNetImg2ImgPipeline(
vae=vae,
controlnet=controlnet_canny,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
force_zeros_for_empty_prompt=False
)
pipe_pose = StableDiffusionXLControlNetImg2ImgPipeline(
vae=vae,
controlnet=controlnet_pose,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
force_zeros_for_empty_prompt=False
)
# ๋ฒˆ์—ญ ๋ชจ๋ธ ์ดˆ๊ธฐํ™”
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")
# prompt๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ํ•จ์ˆ˜ ์ถ”๊ฐ€
def process_prompt(prompt):
if any('\u3131' <= char <= '\u3163' or '\uac00' <= char <= '\ud7a3' for char in prompt):
translated = translator(prompt)[0]['translation_text']
return prompt, translated
return prompt, prompt
@spaces.GPU
def process_canny_condition(image, canny_threods=[100,200]):
np_image = image.copy()
np_image = cv2.Canny(np_image, canny_threods[0], canny_threods[1])
np_image = np_image[:, :, None]
np_image = np.concatenate([np_image, np_image, np_image], axis=2)
np_image = HWC3(np_image)
return Image.fromarray(np_image)
model_midas = MidasDetector()
@spaces.GPU
def process_depth_condition_midas(img, res = 1024):
h,w,_ = img.shape
img = resize_image(HWC3(img), res)
result = HWC3(model_midas(img))
result = cv2.resize(result, (w,h))
return Image.fromarray(result)
model_dwpose = DWposeDetector()
@spaces.GPU
def process_dwpose_condition(image, res=1024):
h,w,_ = image.shape
img = resize_image(HWC3(image), res)
out_res, out_img = model_dwpose(image)
result = HWC3(out_img)
result = cv2.resize(result, (w,h))
return Image.fromarray(result)
MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024
@spaces.GPU
def infer_depth(prompt,
image = None,
negative_prompt = "NSFW, facial shadow, low resolution, JPEG artifacts, blurry, poor quality, blackface, neon lights.",
seed = 397886929,
randomize_seed = False,
guidance_scale = 6.0,
num_inference_steps = 50,
controlnet_conditioning_scale = 0.7,
control_guidance_end = 0.9,
strength = 1.0
):
original_prompt, english_prompt = process_prompt(prompt)
if randomize_seed:
seed = random.randint(0, MAX_SEED)
generator = torch.Generator().manual_seed(seed)
init_image = resize_image(image, MAX_IMAGE_SIZE)
pipe = pipe_depth.to("cuda")
condi_img = process_depth_condition_midas(np.array(init_image), MAX_IMAGE_SIZE)
image = pipe(
prompt=english_prompt,
image=init_image,
controlnet_conditioning_scale=controlnet_conditioning_scale,
control_guidance_end=control_guidance_end,
strength=strength,
control_image=condi_img,
negative_prompt=negative_prompt,
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
num_images_per_prompt=1,
generator=generator,
).images[0]
return [condi_img, image], seed, original_prompt, english_prompt
@spaces.GPU
def infer_canny(prompt,
image = None,
negative_prompt = "NSFW, facial shadow, low resolution, JPEG artifacts, blurry, poor quality, blackface, neon lights.",
seed = 397886929,
randomize_seed = False,
guidance_scale = 6.0,
num_inference_steps = 50,
controlnet_conditioning_scale = 0.7,
control_guidance_end = 0.9,
strength = 1.0
):
original_prompt, english_prompt = process_prompt(prompt)
if randomize_seed:
seed = random.randint(0, MAX_SEED)
generator = torch.Generator().manual_seed(seed)
init_image = resize_image(image, MAX_IMAGE_SIZE)
pipe = pipe_canny.to("cuda")
condi_img = process_canny_condition(np.array(init_image))
image = pipe(
prompt=english_prompt,
image=init_image,
controlnet_conditioning_scale=controlnet_conditioning_scale,
control_guidance_end=control_guidance_end,
strength=strength,
control_image=condi_img,
negative_prompt=negative_prompt,
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
num_images_per_prompt=1,
generator=generator,
).images[0]
return [condi_img, image], seed, original_prompt, english_prompt
@spaces.GPU
def infer_pose(prompt,
image = None,
negative_prompt = "NSFW, facial shadow, low resolution, JPEG artifacts, blurry, poor quality, blackface, neon lights.",
seed = 66,
randomize_seed = False,
guidance_scale = 6.0,
num_inference_steps = 50,
controlnet_conditioning_scale = 0.7,
control_guidance_end = 0.9,
strength = 1.0
):
original_prompt, english_prompt = process_prompt(prompt)
if randomize_seed:
seed = random.randint(0, MAX_SEED)
generator = torch.Generator().manual_seed(seed)
init_image = resize_image(image, MAX_IMAGE_SIZE)
pipe = pipe_pose.to("cuda")
condi_img = process_dwpose_condition(np.array(init_image), MAX_IMAGE_SIZE)
image = pipe(
prompt=english_prompt,
image=init_image,
controlnet_conditioning_scale=controlnet_conditioning_scale,
control_guidance_end=control_guidance_end,
strength=strength,
control_image=condi_img,
negative_prompt=negative_prompt,
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
num_images_per_prompt=1,
generator=generator,
).images[0]
return [condi_img, image], seed, original_prompt, english_prompt
canny_examples = [
["์•„๋ฆ„๋‹ค์šด ์†Œ๋…€, ๊ณ ํ’ˆ์งˆ, ๋งค์šฐ ์„ ๋ช…, ์ƒ์ƒํ•œ ์ƒ‰์ƒ, ์ดˆ๊ณ ํ•ด์ƒ๋„, ์ตœ์ƒ์˜ ํ’ˆ์งˆ, 8k, ๊ณ ํ™”์งˆ, 4K",
"image/woman_1.png"],
["ํŒŒ๋…ธ๋ผ๋งˆ, ์ปต ์•ˆ์— ์•‰์•„์žˆ๋Š” ๊ท€์—ฌ์šด ํฐ ๊ฐ•์•„์ง€, ์นด๋ฉ”๋ผ๋ฅผ ๋ฐ”๋ผ๋ณด๋Š”, ์• ๋‹ˆ๋ฉ”์ด์…˜ ์Šคํƒ€์ผ, 3D ๋ Œ๋”๋ง, ์˜ฅํ…Œ์ธ ๋ Œ๋”",
"image/dog.png"]
]
depth_examples = [
["์‹ ์นด์ด ๋งˆ์ฝ”ํ†  ์Šคํƒ€์ผ, ํ’๋ถ€ํ•œ ์ƒ‰๊ฐ, ์ดˆ๋ก ์…”์ธ ๋ฅผ ์ž…์€ ์—ฌ์„ฑ์ด ๋“คํŒ์— ์„œ ์žˆ๋Š”, ์•„๋ฆ„๋‹ค์šด ํ’๊ฒฝ, ๋ง‘๊ณ  ๋ฐ์€, ์–ผ๋ฃฉ์ง„ ๋น›๊ณผ ๊ทธ๋ฆผ์ž, ์ตœ๊ณ ์˜ ํ’ˆ์งˆ, ์ดˆ์„ธ๋ฐ€, 8K ํ™”์งˆ",
"image/woman_2.png"],
["ํ™”๋ คํ•œ ์ƒ‰์ƒ์˜ ์ž‘์€ ์ƒˆ, ๊ณ ํ’ˆ์งˆ, ๋งค์šฐ ์„ ๋ช…, ์ƒ์ƒํ•œ ์ƒ‰์ƒ, ์ดˆ๊ณ ํ•ด์ƒ๋„, ์ตœ์ƒ์˜ ํ’ˆ์งˆ, 8k, ๊ณ ํ™”์งˆ, 4K",
"image/bird.png"]
]
pose_examples = [
["๋ณด๋ผ์ƒ‰ ํผํ”„ ์Šฌ๋ฆฌ๋ธŒ ๋“œ๋ ˆ์Šค๋ฅผ ์ž…๊ณ  ์™•๊ด€๊ณผ ํฐ์ƒ‰ ๋ ˆ์ด์Šค ์žฅ๊ฐ‘์„ ๋‚€ ์†Œ๋…€๊ฐ€ ์–‘ ์†์œผ๋กœ ์–ผ๊ตด์„ ๊ฐ์‹ธ๊ณ  ์žˆ๋Š”, ๊ณ ํ’ˆ์งˆ, ๋งค์šฐ ์„ ๋ช…, ์ƒ์ƒํ•œ ์ƒ‰์ƒ, ์ดˆ๊ณ ํ•ด์ƒ๋„, ์ตœ์ƒ์˜ ํ’ˆ์งˆ, 8k, ๊ณ ํ™”์งˆ, 4K",
"image/woman_3.png"],
["๊ฒ€์€์ƒ‰ ์Šคํฌ์ธ  ์žฌํ‚ท๊ณผ ํฐ์ƒ‰ ์ด๋„ˆ๋ฅผ ์ž…๊ณ  ๋ชฉ๊ฑธ์ด๋ฅผ ํ•œ ์—ฌ์„ฑ์ด ๊ฑฐ๋ฆฌ์— ์„œ ์žˆ๋Š”, ๋ฐฐ๊ฒฝ์€ ๋นจ๊ฐ„ ๊ฑด๋ฌผ๊ณผ ๋…น์ƒ‰ ๋‚˜๋ฌด, ๊ณ ํ’ˆ์งˆ, ๋งค์šฐ ์„ ๋ช…, ์ƒ์ƒํ•œ ์ƒ‰์ƒ, ์ดˆ๊ณ ํ•ด์ƒ๋„, ์ตœ์ƒ์˜ ํ’ˆ์งˆ, 8k, ๊ณ ํ™”์งˆ, 4K",
"image/woman_4.png"]
]
css = """
footer {
visibility: hidden;
}
"""
def load_description(fp):
with open(fp, 'r', encoding='utf-8') as f:
content = f.read()
return content
with gr.Blocks(theme="Nymbo/Nymbo_Theme", css=css) as Kolors:
with gr.Row():
with gr.Column(elem_id="col-left"):
with gr.Row():
prompt = gr.Textbox(
label="ํ”„๋กฌํ”„ํŠธ",
placeholder="ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ž…๋ ฅํ•˜์„ธ์š” (ํ•œ๊ธ€ ๋˜๋Š” ์˜์–ด)",
lines=2
)
with gr.Row():
image = gr.Image(label="์ด๋ฏธ์ง€", type="pil")
with gr.Accordion("๊ณ ๊ธ‰ ์„ค์ •", open=False):
negative_prompt = gr.Textbox(
label="๋„ค๊ฑฐํ‹ฐ๋ธŒ ํ”„๋กฌํ”„ํŠธ",
placeholder="๋„ค๊ฑฐํ‹ฐ๋ธŒ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ž…๋ ฅํ•˜์„ธ์š”",
visible=True,
value="nsfw, ์–ผ๊ตด ๊ทธ๋ฆผ์ž, ์ €ํ•ด์ƒ๋„, jpeg ์•„ํ‹ฐํŒฉํŠธ, ํ๋ฆฟํ•จ, ์—ด์•…ํ•จ, ๊ฒ€์€ ์–ผ๊ตด, ๋„ค์˜จ ์กฐ๋ช…"
)
seed = gr.Slider(
label="์‹œ๋“œ",
minimum=0,
maximum=MAX_SEED,
step=1,
value=0,
)
randomize_seed = gr.Checkbox(label="์‹œ๋“œ ๋ฌด์ž‘์œ„ํ™”", value=True)
with gr.Row():
guidance_scale = gr.Slider(
label="๊ฐ€์ด๋˜์Šค ์Šค์ผ€์ผ",
minimum=0.0,
maximum=10.0,
step=0.1,
value=6.0,
)
num_inference_steps = gr.Slider(
label="์ถ”๋ก  ๋‹จ๊ณ„ ์ˆ˜",
minimum=10,
maximum=50,
step=1,
value=30,
)
with gr.Row():
controlnet_conditioning_scale = gr.Slider(
label="์ปจํŠธ๋กค๋„ท ์ปจ๋””์…”๋‹ ์Šค์ผ€์ผ",
minimum=0.0,
maximum=1.0,
step=0.1,
value=0.7,
)
control_guidance_end = gr.Slider(
label="์ปจํŠธ๋กค ๊ฐ€์ด๋˜์Šค ์ข…๋ฃŒ",
minimum=0.0,
maximum=1.0,
step=0.1,
value=0.9,
)
with gr.Row():
strength = gr.Slider(
label="๊ฐ•๋„",
minimum=0.0,
maximum=1.0,
step=0.1,
value=1.0,
)
with gr.Row():
canny_button = gr.Button("์บ๋‹ˆ", elem_id="button")
depth_button = gr.Button("๊นŠ์ด", elem_id="button")
pose_button = gr.Button("ํฌ์ฆˆ", elem_id="button")
with gr.Column(elem_id="col-right"):
result = gr.Gallery(label="๊ฒฐ๊ณผ", show_label=False, columns=2)
seed_used = gr.Number(label="์‚ฌ์šฉ๋œ ์‹œ๋“œ")
original_prompt_display = gr.Textbox(label="์›๋ณธ ํ”„๋กฌํ”„ํŠธ")
english_prompt_display = gr.Textbox(label="์˜์–ด ํ”„๋กฌํ”„ํŠธ")
with gr.Row():
gr.Examples(
fn=infer_canny,
examples=canny_examples,
inputs=[prompt, image],
outputs=[result, seed_used, original_prompt_display, english_prompt_display],
label="Canny"
)
with gr.Row():
gr.Examples(
fn=infer_depth,
examples=depth_examples,
inputs=[prompt, image],
outputs=[result, seed_used, original_prompt_display, english_prompt_display],
label="Depth"
)
with gr.Row():
gr.Examples(
fn=infer_pose,
examples=pose_examples,
inputs=[prompt, image],
outputs=[result, seed_used, original_prompt_display, english_prompt_display],
label="Pose"
)
canny_button.click(
fn=infer_canny,
inputs=[prompt, image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps, controlnet_conditioning_scale, control_guidance_end, strength],
outputs=[result, seed_used, original_prompt_display, english_prompt_display]
)
depth_button.click(
fn=infer_depth,
inputs=[prompt, image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps, controlnet_conditioning_scale, control_guidance_end, strength],
outputs=[result, seed_used, original_prompt_display, english_prompt_display]
)
pose_button.click(
fn=infer_pose,
inputs=[prompt, image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps, controlnet_conditioning_scale, control_guidance_end, strength],
outputs=[result, seed_used, original_prompt_display, english_prompt_display]
)
Kolors.queue().launch(debug=True)