app.py

import spaces
import random
import torch
from huggingface_hub import snapshot_download
from transformers import CLIPVisionModelWithProjection, CLIPImageProcessor
from kolors.pipelines import pipeline_stable_diffusion_xl_chatglm_256_ipadapter, pipeline_stable_diffusion_xl_chatglm_256
from kolors.models.modeling_chatglm import ChatGLMModel
from kolors.models.tokenization_chatglm import ChatGLMTokenizer
from kolors.models import unet_2d_condition
from diffusers import AutoencoderKL, EulerDiscreteScheduler, UNet2DConditionModel
import gradio as gr
import numpy as np
from huggingface_hub import InferenceClient
import os
from PIL import Image
import re

device = "cuda"
ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors")
ckpt_IPA_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors-IP-Adapter-Plus")

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_t2i = UNet2DConditionModel.from_pretrained(f"{ckpt_dir}/unet", revision=None).half().to(device)
unet_i2i = unet_2d_condition.UNet2DConditionModel.from_pretrained(f"{ckpt_dir}/unet", revision=None).half().to(device)
image_encoder = CLIPVisionModelWithProjection.from_pretrained(f'{ckpt_IPA_dir}/image_encoder',ignore_mismatched_sizes=True).to(dtype=torch.float16, device=device)
ip_img_size = 336
clip_image_processor = CLIPImageProcessor(size=ip_img_size, crop_size=ip_img_size)

pipe_t2i = pipeline_stable_diffusion_xl_chatglm_256.StableDiffusionXLPipeline(
    vae=vae,
    text_encoder=text_encoder, 
    tokenizer=tokenizer, 
    unet=unet_t2i, 
    scheduler=scheduler, 
    force_zeros_for_empty_prompt=False
).to(device)

pipe_i2i = pipeline_stable_diffusion_xl_chatglm_256_ipadapter.StableDiffusionXLPipeline(
    vae=vae,
    text_encoder=text_encoder,
    tokenizer=tokenizer,
    unet=unet_i2i,
    scheduler=scheduler,
    image_encoder=image_encoder,
    feature_extractor=clip_image_processor,
    force_zeros_for_empty_prompt=False
).to(device)

if hasattr(pipe_i2i.unet, 'encoder_hid_proj'):
    pipe_i2i.unet.text_encoder_hid_proj = pipe_i2i.unet.encoder_hid_proj
    
pipe_i2i.load_ip_adapter(f'{ckpt_IPA_dir}' , subfolder="", weight_name=["ip_adapter_plus_general.bin"])

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024

@spaces.GPU
def infer(prompt, 
          ip_adapter_image = None, 
          ip_adapter_scale = 0.5, 
          negative_prompt = "", 
          seed = 0, 
          randomize_seed = False, 
          width = 1024, 
          height = 1024, 
          guidance_scale = 5.0, 
          num_inference_steps = 25
          ):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    generator = torch.Generator().manual_seed(seed)

    if ip_adapter_image is None:
        pipe_t2i.to(device)
        image = pipe_t2i(
            prompt = prompt, 
            negative_prompt = negative_prompt,
            guidance_scale = guidance_scale, 
            num_inference_steps = num_inference_steps, 
            width = width, 
            height = height,
            generator = generator
        ).images[0] 
        image.save("generated_image.jpg")  # 파일 확장자를 .jpg로 변경
        return image, "generated_image.jpg"
    else:
        pipe_i2i.to(device)
        image_encoder.to(device)
        pipe_i2i.image_encoder = image_encoder
        pipe_i2i.set_ip_adapter_scale([ip_adapter_scale])
        image = pipe_i2i(
                prompt=prompt,
                ip_adapter_image=[ip_adapter_image],
                negative_prompt=negative_prompt, 
                height=height,
                width=width,
                num_inference_steps=num_inference_steps, 
                guidance_scale=guidance_scale,
                num_images_per_prompt=1,
                generator=generator
            ).images[0]
        image.save("generated_image.jpg")  # 파일 확장자를 .jpg로 변경
        return image, "generated_image.jpg"

css="""
#col-left {
    margin: 0 auto;
    max-width: 600px;
}
#col-right {
    margin: 0 auto;
    max-width: 750px;
}
#output {
    height: 500px; 
    overflow: auto; 
    border: 1px solid #ccc; 
}
"""

# 추가 코드 통합
from transformers import AutoProcessor, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained('gokaygokay/Florence-2-SD3-Captioner', trust_remote_code=True).eval()
processor = AutoProcessor.from_pretrained('gokaygokay/Florence-2-SD3-Captioner', trust_remote_code=True)

def modify_caption(caption: str) -> str:
    prefix_substrings = [
        ('captured from ', ''),
        ('captured at ', '')
    ]
    pattern = '|'.join([re.escape(opening) for opening, _ in prefix_substrings])
    replacers = {opening.lower(): replacer for opening, replacer in prefix_substrings}
    def replace_fn(match):
        return replacers[match.group(0).lower()]
    modified_caption = re.sub(pattern, replace_fn, caption, count=1, flags=re.IGNORECASE)
    return modified_caption if modified_caption != caption else caption
    
@spaces.GPU
def run_example(image):
    image = Image.fromarray(image)
    task_prompt = "<DESCRIPTION>"
    prompt = task_prompt + "Describe this image in great detail."

    if image.mode != "RGB":
        image = image.convert("RGB")

    inputs = processor(text=prompt, images=image, return_tensors="pt")
    generated_ids = model.generate(
        input_ids=inputs["input_ids"],
        pixel_values=inputs["pixel_values"],
        max_new_tokens=1024,
        num_beams=3
    )
    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
    parsed_answer = processor.post_process_generation(generated_text, task=task_prompt, image_size=(image.width, image.height))
    return modify_caption(parsed_answer["<DESCRIPTION>"])

with gr.Blocks(css=css) as Kolors:
    with gr.Tab("Image Generation"):
        with gr.Row():
            with gr.Column(elem_id="col-left"):
                with gr.Row():
                    generated_prompt = gr.Textbox(
                        label="프롬프트 입력",
                        placeholder="이미지 생성에 사용할 프롬프트를 입력하세요",
                        lines=2
                    )
                with gr.Row():
                    ip_adapter_image = gr.Image(label="Image Prompt (optional)", type="pil")
                with gr.Row(visible=False):  # Advanced Settings 숨김
                    negative_prompt = gr.Textbox(
                        label="Negative prompt",
                        placeholder="Enter a negative prompt",
                        visible=True,
                    )
                    seed = gr.Slider(
                        label="Seed",
                        minimum=0,
                        maximum=MAX_SEED,
                        step=1,
                        value=0,
                    )
                    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=1024,
                        )
                    with gr.Row():
                        guidance_scale = gr.Slider(
                            label="Guidance scale",
                            minimum=0.0,
                            maximum=10.0,
                            step=0.1,
                            value=5.0,
                        )
                        num_inference_steps = gr.Slider(
                            label="Number of inference steps",
                            minimum=10,
                            maximum=50,
                            step=1,
                            value=25,
                        )
                    with gr.Row():
                        ip_adapter_scale = gr.Slider(
                            label="Image influence scale",
                            info="Use 1 for creating variations",
                            minimum=0.0,
                            maximum=1.0,
                            step=0.05,
                            value=0.5,
                        )
                with gr.Row():
                    run_button = gr.Button("Generate Image")
                
            with gr.Column(elem_id="col-right"):
                result = gr.Image(label="Result", show_label=False)
                download_button = gr.File(label="Download Image")

        # 이미지 생성 및 다운로드 파일 경로 설정
        run_button.click(
            fn=infer,
            inputs=[generated_prompt, ip_adapter_image, ip_adapter_scale, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
            outputs=[result, download_button]
        )
    
    with gr.Tab("Florence-2 SD3 Prompts"):
        gr.Markdown("# [Florence-2 SD3 Long Captioner](https://huggingface.co/gokaygokay/Florence-2-SD3-Captioner/)")
        gr.Markdown("[Florence-2 Base](https://huggingface.co/microsoft/Florence-2-base-ft) fine-tuned on Long SD3 Prompt and Image pairs. Check above link for datasets that are used for fine-tuning.")
        with gr.Row():
            with gr.Column():
                input_img = gr.Image(label="Input Picture")
                submit_btn = gr.Button(value="Submit")
            with gr.Column():
                output_text = gr.Textbox(label="Output Text")


        submit_btn.click(run_example, [input_img], [output_text])

Kolors.queue().launch(debug=True)