chocs / app.py
seawolf2357's picture
Update app.py
2caff40 verified
raw
history blame
10.9 kB
import random
import os
import uuid
import re
import time
from datetime import datetime
import gradio as gr
import numpy as np
import requests
import torch
from diffusers import DiffusionPipeline
from PIL import Image
# ===== OpenAI ์„ค์ • =====
from openai import OpenAI
client = OpenAI(api_key=os.getenv("LLM_API")) # ํ™˜๊ฒฝ ๋ณ€์ˆ˜์— API ํ‚ค๊ฐ€ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
# ===== ํ”„๋กฌํ”„ํŠธ ์ฆ๊ฐ•์šฉ ์Šคํƒ€์ผ ํ”„๋ฆฌ์…‹ =====
STYLE_PRESETS = {
"None": "",
"Realistic Photo": "photorealistic, 8k, ultra-detailed, cinematic lighting, realistic skin texture",
"Oil Painting": "oil painting, rich brush strokes, canvas texture, baroque lighting",
"Comic Book": "comic book style, bold ink outlines, cel shading, vibrant colors",
"Watercolor": "watercolor illustration, soft gradients, splatter effect, pastel palette",
}
# ===== ์ €์žฅ ํด๋” =====
SAVE_DIR = "saved_images" # Gradio will handle the persistence
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR, exist_ok=True)
# ===== ๋””๋ฐ”์ด์Šค & ๋ชจ๋ธ ๋กœ๋“œ =====
device = "cuda" if torch.cuda.is_available() else "cpu"
repo_id = "black-forest-labs/FLUX.1-dev"
adapter_id = "seawolf2357/kim-korea" # ํŠน์ • ์ •์น˜์ธ์„ ํ•™์Šตํ•œ LoRA ๋ชจ๋ธ
pipeline = DiffusionPipeline.from_pretrained(repo_id, torch_dtype=torch.bfloat16)
pipeline.load_lora_weights(adapter_id)
pipeline = pipeline.to(device)
MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024
# ===== ํ•œ๊ธ€ ์—ฌ๋ถ€ ํŒ๋ณ„ =====
HANGUL_RE = re.compile(r"[\u3131-\u318E\uAC00-\uD7A3]+")
def is_korean(text: str) -> bool:
return bool(HANGUL_RE.search(text))
# ===== ๋ฒˆ์—ญ & ์ฆ๊ฐ• ํ•จ์ˆ˜ =====
def openai_translate(text: str, retries: int = 3) -> str:
"""ํ•œ๊ธ€์„ ์˜์–ด๋กœ ๋ฒˆ์—ญ (OpenAI GPT-4.1-mini ์‚ฌ์šฉ). ์˜์–ด ์ž…๋ ฅ์ด๋ฉด ๊ทธ๋Œ€๋กœ ๋ฐ˜ํ™˜."""
if not is_korean(text):
return text
for attempt in range(retries):
try:
res = client.chat.completions.create(
model="gpt-4.1-mini",
messages=[
{
"role": "system",
"content": "Translate the following Korean prompt into concise, descriptive English suitable for an image generation model. Keep the meaning, do not add new concepts."
},
{"role": "user", "content": text}
],
temperature=0.3,
max_tokens=256,
)
return res.choices[0].message.content.strip()
except (requests.exceptions.RequestException, Exception) as e:
print(f"[translate] attempt {attempt + 1} failed: {e}")
time.sleep(2)
return text # ๋ฒˆ์—ญ ์‹คํŒจ ์‹œ ์›๋ฌธ ๊ทธ๋Œ€๋กœ
def prepare_prompt(user_prompt: str, style_key: str) -> str:
"""ํ•œ๊ธ€์ด๋ฉด ๋ฒˆ์—ญํ•˜๊ณ , ์„ ํƒํ•œ ์Šคํƒ€์ผ ํ”„๋ฆฌ์…‹์„ ๋ถ™์—ฌ์„œ ์ตœ์ข… ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋งŒ๋“ ๋‹ค."""
prompt_en = openai_translate(user_prompt)
style_suffix = STYLE_PRESETS.get(style_key, "")
if style_suffix:
final_prompt = f"{prompt_en}, {style_suffix}"
else:
final_prompt = prompt_en
return final_prompt
# ===== ์ด๋ฏธ์ง€ ์ €์žฅ =====
def save_generated_image(image: Image.Image, prompt: str) -> str:
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
unique_id = str(uuid.uuid4())[:8]
filename = f"{timestamp}_{unique_id}.png"
filepath = os.path.join(SAVE_DIR, filename)
image.save(filepath)
# ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ์ €์žฅ
metadata_file = os.path.join(SAVE_DIR, "metadata.txt")
with open(metadata_file, "a", encoding="utf-8") as f:
f.write(f"{filename}|{prompt}|{timestamp}\n")
return filepath
# ===== Diffusion ํ˜ธ์ถœ =====
def run_pipeline(prompt: str, seed: int, width: int, height: int, guidance_scale: float, num_steps: int, lora_scale: float):
generator = torch.Generator(device=device).manual_seed(int(seed))
result = pipeline(
prompt=prompt,
guidance_scale=guidance_scale,
num_inference_steps=num_steps,
width=width,
height=height,
generator=generator,
joint_attention_kwargs={"scale": lora_scale},
).images[0]
return result
# ===== Gradio inference ๋ž˜ํผ =====
@spaces.GPU(duration=60)
def generate_image(
user_prompt: str,
style_key: str,
seed: int = 42,
randomize_seed: bool = True,
width: int = 1024,
height: int = 768,
guidance_scale: float = 3.5,
num_inference_steps: int = 30,
lora_scale: float = 1.0,
progress=None,
):
if randomize_seed:
seed = random.randint(0, MAX_SEED)
# 1) ๋ฒˆ์—ญ + ์ฆ๊ฐ•
final_prompt = prepare_prompt(user_prompt, style_key)
# 2) ํŒŒ์ดํ”„๋ผ์ธ ํ˜ธ์ถœ
image = run_pipeline(final_prompt, seed, width, height, guidance_scale, num_inference_steps, lora_scale)
# 3) ์ €์žฅ
save_generated_image(image, final_prompt)
return image, seed
# ===== ์˜ˆ์‹œ ํ”„๋กฌํ”„ํŠธ (ํ•œ๊ตญ์–ด/์˜์–ด ํ˜ผ์šฉ ํ—ˆ์šฉ) =====
examples = [
"Mr. KIM์ด ๋‘ ์†์œผ๋กœ 'Fighting!' ํ˜„์ˆ˜๋ง‰์„ ๋“ค๊ณ  ์žˆ๋Š” ๋ชจ์Šต, ์• ๊ตญ์‹ฌ๊ณผ ๊ตญ๊ฐ€ ๋ฐœ์ „์— ๋Œ€ํ•œ ์˜์ง€๋ฅผ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ๋‹ค.",
"Mr. KIM์ด ์–‘ํŒ”์„ ๋“ค์–ด ์˜ฌ๋ฆฌ๋ฉฐ ์Šน๋ฆฌ์˜ ํ‘œ์ •์œผ๋กœ ํ™˜ํ˜ธํ•˜๋Š” ๋ชจ์Šต, ์Šน๋ฆฌ์™€ ๋ฏธ๋ž˜์— ๋Œ€ํ•œ ํฌ๋ง์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ๋‹ค.",
"Mr. KIM์ด ์šด๋™๋ณต์„ ์ž…๊ณ  ๊ณต์›์—์„œ ์กฐ๊น…ํ•˜๋Š” ๋ชจ์Šต, ๊ฑด๊ฐ•ํ•œ ์ƒํ™œ์Šต๊ด€๊ณผ ํ™œ๊ธฐ์ฐฌ ๋ฆฌ๋”์‹ญ์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ๋‹ค.",
"Mr. KIM์ด ๋ถ๋น„๋Š” ๊ฑฐ๋ฆฌ์—์„œ ์—ฌ์„ฑ ์‹œ๋ฏผ๋“ค๊ณผ ๋”ฐ๋œปํ•˜๊ฒŒ ์•…์ˆ˜ํ•˜๋Š” ๋ชจ์Šต, ์—ฌ์„ฑ ์œ ๊ถŒ์ž๋“ค์— ๋Œ€ํ•œ ์ง„์ •ํ•œ ๊ด€์‹ฌ๊ณผ ์†Œํ†ต์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ๋‹ค.",
"Mr. KIM์ด ์„ ๊ฑฐ ์œ ์„ธ์žฅ์—์„œ ์ง€ํ‰์„ ์„ ํ–ฅํ•ด ์†๊ฐ€๋ฝ์œผ๋กœ ๊ฐ€๋ฆฌํ‚ค๋ฉฐ ์˜๊ฐ์„ ์ฃผ๋Š” ์ œ์Šค์ฒ˜๋ฅผ ์ทจํ•˜๊ณ  ์žˆ๊ณ , ์—ฌ์„ฑ๋“ค๊ณผ ์•„์ด๋“ค์ด ๋ฐ•์ˆ˜๋ฅผ ์น˜๊ณ  ์žˆ๋‹ค.",
"Mr. KIM์ด ์ง€์—ญ ํ–‰์‚ฌ์— ์ฐธ์—ฌํ•˜์—ฌ ์—ด์ •์ ์œผ๋กœ ์‘์›ํ•˜๋Š” ์—ฌ์„ฑ ์ง€์ง€์ž๋“ค์—๊ฒŒ ๋‘˜๋Ÿฌ์‹ธ์—ฌ ์žˆ๋Š” ๋ชจ์Šต.",
"Mr. KIM visiting a local market, engaging in friendly conversation with female vendors and shopkeepers. ",
"Mr. KIM walking through a university campus, discussing education policies with female students and professors. ",
"Mr. KIM delivering a powerful speech in front of a large crowd with confident gestures and determined expression. ",
"Mr. KIM in a dynamic interview setting, passionately outlining his visions for the future.",
"Mr. KIM preparing for an important debate, surrounded by paperwork, looking focused and resolute. ",
]
# ===== ์ปค์Šคํ…€ CSS (๋ถ‰์€ ํ†ค ์œ ์ง€) =====
custom_css = """
:root {
--color-primary: #8F1A3A;
--color-secondary: #FF4B4B;
--background-fill-primary: linear-gradient(to right, #FFF5F5, #FED7D7, #FEB2B2);
}
footer {visibility: hidden;}
.gradio-container {background: var(--background-fill-primary);}
.title {color: var(--color-primary)!important; font-size:3rem!important; font-weight:700!important; text-align:center; margin:1rem 0; font-family:'Playfair Display',serif;}
.subtitle {color:#4A5568!important; font-size:1.2rem!important; text-align:center; margin-bottom:1.5rem; font-style:italic;}
.collection-link {text-align:center; margin-bottom:2rem; font-size:1.1rem;}
.collection-link a {color:var(--color-primary); text-decoration:underline; transition:color .3s ease;}
.collection-link a:hover {color:var(--color-secondary);}
.model-description{background:rgba(255,255,255,.8); border-radius:12px; padding:24px; margin:20px 0; box-shadow:0 4px 12px rgba(0,0,0,.05); border-left:5px solid var(--color-primary);}
button.primary{background:var(--color-primary)!important; color:#fff!important; transition:all .3s ease;}
button:hover{transform:translateY(-2px); box-shadow:0 5px 15px rgba(0,0,0,.1);}
.input-container{border-radius:10px; box-shadow:0 2px 8px rgba(0,0,0,.05); background:rgba(255,255,255,.6); padding:20px; margin-bottom:1rem;}
.advanced-settings{margin-top:1rem; padding:1rem; border-radius:10px; background:rgba(255,255,255,.6);}
.example-region{background:rgba(255,255,255,.5); border-radius:10px; padding:1rem; margin-top:1rem;}
"""
# ===== Gradio UI =====
with gr.Blocks(css=custom_css, analytics_enabled=False) as demo:
gr.HTML('<div class="title">Mr. KIM in KOREA</div>')
gr.HTML('<div class="collection-link"><a href="https://huggingface.co/collections/openfree/painting-art-ai-681453484ec15ef5978bbeb1" target="_blank">Visit the LoRA Model Collection</a></div>')
with gr.Group(elem_classes="model-description"):
gr.HTML("""
<p>
๋ณธ ๋ชจ๋ธ์€ ์—ฐ๊ตฌ ๋ชฉ์ ์œผ๋กœ ํŠน์ •์ธ์˜ ์–ผ๊ตด๊ณผ ์™ธ๋ชจ๋ฅผ ํ•™์Šตํ•œ LoRA ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.<br>
๋ชฉ์ ์™ธ์˜ ์šฉ๋„๋กœ ๋ฌด๋‹จ ์‚ฌ์šฉ ์•Š๋„๋ก ์œ ์˜ํ•ด ์ฃผ์„ธ์š”.<br>
(์˜ˆ์‹œ prompt ์‚ฌ์šฉ ์‹œ ๋ฐ˜๋“œ์‹œ 'kim'์„ ํฌํ•จํ•˜์—ฌ์•ผ ์ตœ์ ์˜ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.)
</p>
""")
# ===== ๋ฉ”์ธ ์ž…๋ ฅ =====
with gr.Column():
with gr.Row(elem_classes="input-container"):
user_prompt = gr.Text(label="Prompt", max_lines=1, value=examples[0])
style_select = gr.Radio(label="Style Preset", choices=list(STYLE_PRESETS.keys()), value="None", interactive=True)
run_button = gr.Button("Generate", variant="primary")
result_image = gr.Image(label="Generated Image")
seed_output = gr.Number(label="Seed")
# ===== ๊ณ ๊ธ‰ ์„ค์ • =====
with gr.Accordion("Advanced Settings", open=False, elem_classes="advanced-settings"):
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42)
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
with gr.Row():
width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=768)
with gr.Row():
guidance_scale = gr.Slider(label="Guidance scale", minimum=0.0, maximum=10.0, step=0.1, value=3.5)
num_inference_steps = gr.Slider(label="Inference steps", minimum=1, maximum=50, step=1, value=30)
lora_scale = gr.Slider(label="LoRA scale", minimum=0.0, maximum=1.0, step=0.1, value=1.0)
# ===== ์˜ˆ์‹œ ์˜์—ญ =====
with gr.Group(elem_classes="example-region"):
gr.Markdown("### Examples")
gr.Examples(examples=examples, inputs=user_prompt, cache_examples=False)
# ===== ์ด๋ฒคํŠธ =====
run_button.click(
fn=generate_image,
inputs=[
user_prompt,
style_select,
seed,
randomize_seed,
width,
height,
guidance_scale,
num_inference_steps,
lora_scale,
],
outputs=[result_image, seed_output],
)
demo.queue()
demo.launch()