app.py

import os
from huggingface_hub import snapshot_download

os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"

REVISION = "ceaf371f01ef66192264811b390bccad475a4f02"

LOCAL_FLORENCE = snapshot_download(
    repo_id="microsoft/Florence-2-base",
    revision=REVISION
)

LOCAL_TURBOX = snapshot_download(
    repo_id="tensorart/stable-diffusion-3.5-large-TurboX"
)

LOCAL_FLORENCE_DIR = snapshot_download(
    repo_id="microsoft/Florence-2-base",
    revision=REVISION,
    local_files_only=False
)

import sys, types, importlib.machinery, importlib

spec = importlib.machinery.ModuleSpec('flash_attn', loader=None)
mod = types.ModuleType('flash_attn')
mod.__spec__ = spec
sys.modules['flash_attn'] = mod

import huggingface_hub as _hf_hub
_hf_hub.cached_download = _hf_hub.hf_hub_download

import gradio as gr
import torch
import random
from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM
from transformers import (
    CLIPTextModel,
    CLIPTokenizer,
    CLIPFeatureExtractor,
)
import diffusers
from diffusers import StableDiffusionPipeline
from diffusers import DiffusionPipeline
from diffusers import EulerDiscreteScheduler
from diffusers import UNet2DConditionModel
diffusers.FlowMatchEulerDiscreteScheduler = EulerDiscreteScheduler

import transformers.utils.import_utils as _import_utils
from transformers.utils import is_flash_attn_2_available
_import_utils._is_package_available   = lambda pkg: False
_import_utils.is_flash_attn_2_available = lambda: False

hf_utils = importlib.import_module('transformers.utils')
hf_utils.is_flash_attn_2_available            = lambda *a, **k: False
hf_utils.is_flash_attn_greater_or_equal_2_10 = lambda *a, **k: False

mask_utils = importlib.import_module("transformers.modeling_attn_mask_utils")
for fn in ("_prepare_4d_attention_mask_for_sdpa", "_prepare_4d_causal_attention_mask_for_sdpa"):
    if not hasattr(mask_utils, fn):
        setattr(mask_utils, fn, lambda *a, **k: None)

cfg_mod = importlib.import_module("transformers.configuration_utils")
_PrC = cfg_mod.PretrainedConfig
_orig_getattr = _PrC.__getattribute__
def _getattr(self, name):
    if name == "_attn_implementation":
        return "sdpa"
    return _orig_getattr(self, name)
_PrC.__getattribute__ = _getattr

model_repo = "tensorart/stable-diffusion-3.5-large-TurboX"

# Florence-2 로드
device = "cuda" if torch.cuda.is_available() else "cpu"
scheduler = EulerDiscreteScheduler.from_pretrained(
    model_repo,
    subfolder="scheduler",
    torch_dtype=torch.float16,
)
text_encoder      = CLIPTextModel.from_pretrained(
    model_repo, subfolder="text_encoder_1", torch_dtype=torch.float16
)
tokenizer         = CLIPTokenizer.from_pretrained(
    model_repo, subfolder="tokenizer_1"
)
feature_extractor = CLIPFeatureExtractor.from_pretrained(
    model_repo, subfolder="feature_extractor"
)
unet = UNet2DConditionModel.from_pretrained(
    model_repo, subfolder="unet", torch_dtype=torch.float16
)
florence_model = AutoModelForCausalLM.from_pretrained(LOCAL_FLORENCE, trust_remote_code=True, torch_dtype=torch.float16)
florence_model.to("cpu")
florence_model.eval()
florence_processor = AutoProcessor.from_pretrained(LOCAL_FLORENCE, trust_remote_code=True)

# Stable Diffusion TurboX 로드

diffusers.StableDiffusion3Pipeline = StableDiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "tensorart/stable-diffusion-3.5-large-TurboX",
    torch_dtype=torch.float16,
    trust_remote_code=True,
    safety_checker=None,
    feature_extractor=None
)
pipe = pipe.to("cuda")

pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(model_repo, subfolder="scheduler", local_files_only=True, trust_remote_code = True, shift=5)

MAX_SEED = 2**31 - 1

def pseudo_translate_to_korean_style(en_prompt: str) -> str:
    # 번역 없이 스타일 적용
    return f"Cartoon styled {en_prompt} handsome or pretty people"

def generate_prompt(image):
    """이미지 → 영어 설명 → 한국어 프롬프트 스타일로 변환"""
    if not isinstance(image, Image.Image):
        image = Image.fromarray(image)

    inputs = florence_processor(text="<MORE_DETAILED_CAPTION>", images=image, return_tensors="pt").to(device)
    generated_ids = florence_model.generate(
        input_ids=inputs["input_ids"],
        pixel_values=inputs["pixel_values"],
        max_new_tokens=512,
        num_beams=3
    )
    generated_text = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
    parsed_answer = florence_processor.post_process_generation(
        generated_text,
        task="<MORE_DETAILED_CAPTION>",
        image_size=(image.width, image.height)
    )
    prompt_en = parsed_answer["<MORE_DETAILED_CAPTION>"]

    # 번역기 없이 스타일 적용
    cartoon_prompt = pseudo_translate_to_korean_style(prompt_en)
    return cartoon_prompt

def generate_image(prompt, seed=42, randomize_seed=False):
    """텍스트 프롬프트 → 이미지 생성"""
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    generator = torch.Generator().manual_seed(seed)
    image = pipe(
        prompt=prompt,
        guidance_scale=1.5,
        num_inference_steps=8,
        width=768,
        height=768,
        generator=generator
    ).images[0]
    return image, seed

# Gradio UI 구성
with gr.Blocks() as demo:
    gr.Markdown("# 🖼 이미지 → 설명 생성 → 카툰 이미지 자동 생성기")
    
    gr.Markdown("**📌 사용법 안내 (한국어)**\n"
                "- 왼쪽에 이미지를 업로드하세요.\n"
                "- AI가 영어 설명을 만들고, 내부에서 한국어 스타일 프롬프트로 재구성합니다.\n"
                "- 오른쪽에 결과 이미지가 생성됩니다.")

    with gr.Row():
        with gr.Column():
            input_img = gr.Image(label="🎨 원본 이미지 업로드")
            run_button = gr.Button("✨ 생성 시작")

        with gr.Column():
            prompt_out = gr.Textbox(label="📝 스타일 적용된 프롬프트", lines=3, show_copy_button=True)
            output_img = gr.Image(label="🎉 생성된 이미지")

    def full_process(img):
        prompt = generate_prompt(img)
        image, seed = generate_image(prompt, randomize_seed=True)
        return prompt, image

    run_button.click(fn=full_process, inputs=[input_img], outputs=[prompt_out, output_img])

demo.launch()