app.py

# インポートと初期設定

import spaces
import torch
import gradio as gr
from gradio import processing_utils, utils
from PIL import Image
import random

from diffusers import (
    DiffusionPipeline,
    AutoencoderKL,
    StableDiffusionControlNetPipeline,
    ControlNetModel,
    StableDiffusionLatentUpscalePipeline,
    StableDiffusionImg2ImgPipeline,
    StableDiffusionControlNetImg2ImgPipeline,
    DPMSolverMultistepScheduler,
    EulerDiscreteScheduler
)
import tempfile
import time
from share_btn import community_icon_html, loading_icon_html, share_js
import user_history
from illusion_style import css
import os
from transformers import CLIPImageProcessor
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker



# モデルの初期化

BASE_MODEL = "SG161222/Realistic_Vision_V5.1_noVAE"

# Initialize both pipelines
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16)
controlnet = ControlNetModel.from_pretrained("monster-labs/control_v1p_sd15_qrcode_monster", torch_dtype=torch.float16)

# Initialize the safety checker conditionally
SAFETY_CHECKER_ENABLED = os.environ.get("SAFETY_CHECKER", "0") == "1"
# safety_checker = None
# feature_extractor = None
# if SAFETY_CHECKER_ENABLED:
#     safety_checker = StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker").to("cuda")
#     feature_extractor = CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch32")

# Initialize the safety checker conditionally
SAFETY_CHECKER_ENABLED = False  # 強制的に無効化
safety_checker = None
feature_extractor = None

main_pipe = StableDiffusionControlNetPipeline.from_pretrained(
    BASE_MODEL,
    controlnet=controlnet,
    vae=vae,
    safety_checker=safety_checker,
    feature_extractor=feature_extractor,
    torch_dtype=torch.float16,
).to("cuda")



# 関数の定義

# Function to check NSFW images
#def check_nsfw_images(images: list[Image.Image]) -> tuple[list[Image.Image], list[bool]]:
#    if SAFETY_CHECKER_ENABLED:
#        safety_checker_input = feature_extractor(images, return_tensors="pt").to("cuda")
#        has_nsfw_concepts = safety_checker(
#            images=[images],
#            clip_input=safety_checker_input.pixel_values.to("cuda")
#        )
#        return images, has_nsfw_concepts
#    else:
#        return images, [False] * len(images)
        
#main_pipe.unet = torch.compile(main_pipe.unet, mode="reduce-overhead", fullgraph=True)
#main_pipe.unet.to(memory_format=torch.channels_last)
#main_pipe.unet = torch.compile(main_pipe.unet, mode="reduce-overhead", fullgraph=True)
#model_id = "stabilityai/sd-x2-latent-upscaler"
image_pipe = StableDiffusionControlNetImg2ImgPipeline(**main_pipe.components)


#image_pipe.unet = torch.compile(image_pipe.unet, mode="reduce-overhead", fullgraph=True)
#upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(model_id, torch_dtype=torch.float16)
#upscaler.to("cuda")


# Sampler map
SAMPLER_MAP = {
    "DPM++ Karras SDE": lambda config: DPMSolverMultistepScheduler.from_config(config, use_karras=True, algorithm_type="sde-dpmsolver++"),
    "Euler": lambda config: EulerDiscreteScheduler.from_config(config),
}

# 入力画像を中央からクロップし、指定されたサイズにリサイズする
def center_crop_resize(img, output_size=(512, 512)):
    width, height = img.size

    # Calculate dimensions to crop to the center
    new_dimension = min(width, height)
    left = (width - new_dimension)/2
    top = (height - new_dimension)/2
    right = (width + new_dimension)/2
    bottom = (height + new_dimension)/2

    # Crop and resize
    img = img.crop((left, top, right, bottom))
    img = img.resize(output_size)

    return img

# 指定された方法で画像をアップスケールする
def common_upscale(samples, width, height, upscale_method, crop=False):
        if crop == "center":
            old_width = samples.shape[3]
            old_height = samples.shape[2]
            old_aspect = old_width / old_height
            new_aspect = width / height
            x = 0
            y = 0
            if old_aspect > new_aspect:
                x = round((old_width - old_width * (new_aspect / old_aspect)) / 2)
            elif old_aspect < new_aspect:
                y = round((old_height - old_height * (old_aspect / new_aspect)) / 2)
            s = samples[:,:,y:old_height-y,x:old_width-x]
        else:
            s = samples

        return torch.nn.functional.interpolate(s, size=(height, width), mode=upscale_method)

# common_upscale を利用して画像を指定された倍率でアップスケールする
def upscale(samples, upscale_method, scale_by):
        #s = samples.copy()
        width = round(samples["images"].shape[3] * scale_by)
        height = round(samples["images"].shape[2] * scale_by)
        s = common_upscale(samples["images"], width, height, upscale_method, "disabled")
        return (s)

# ユーザーの入力が適切かどうかをチェックする
def check_inputs(prompt: str, control_image: Image.Image):
    if control_image is None:
        raise gr.Error("Please select or upload an Input Illusion")
    if prompt is None or prompt == "":
        raise gr.Error("Prompt is required")

# Base64エンコードされた画像をPIL（Python Imaging Library）形式の画像に変換する
def convert_to_pil(base64_image):
    pil_image = Image.open(base64_image)
    return pil_image

# PIL形式の画像をBase64形式に変換する
def convert_to_base64(pil_image):
    with tempfile.NamedTemporaryFile(suffix='.png', delete=False) as temp_file:
        image.save(temp_file.name)
    return temp_file.name



# 推論関数
# Inference function

@spaces.GPU
def inference(
    control_image: Image.Image,
    prompt: str,
    negative_prompt: str,
    guidance_scale: float = 8.0,
    controlnet_conditioning_scale: float = 1,
    control_guidance_start: float = 1,    
    control_guidance_end: float = 1,
    upscaler_strength: float = 0.5,
    seed: int = -1,
    sampler = "DPM++ Karras SDE",
    progress = gr.Progress(track_tqdm=True),
    profile: gr.OAuthProfile | None = None,
):
    start_time = time.time()
    start_time_struct = time.localtime(start_time)
    start_time_formatted = time.strftime("%H:%M:%S", start_time_struct)
    print(f"Inference started at {start_time_formatted}")
    
    # Generate the initial image
    #init_image = init_pipe(prompt).images[0]

    # Rest of your existing code
    control_image_small = center_crop_resize(control_image)
    control_image_large = center_crop_resize(control_image, (1024, 1024))

    main_pipe.scheduler = SAMPLER_MAP[sampler](main_pipe.scheduler.config)
    my_seed = random.randint(0, 2**32 - 1) if seed == -1 else seed
    generator = torch.Generator(device="cuda").manual_seed(my_seed)
    
    out = main_pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        image=control_image_small,
        guidance_scale=float(guidance_scale),
        controlnet_conditioning_scale=float(controlnet_conditioning_scale),
        generator=generator,
        control_guidance_start=float(control_guidance_start),
        control_guidance_end=float(control_guidance_end),
        num_inference_steps=15,
        output_type="latent"
    )
    upscaled_latents = upscale(out, "nearest-exact", 2)
    out_image = image_pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        control_image=control_image_large,        
        image=upscaled_latents,
        guidance_scale=float(guidance_scale),
        generator=generator,
        num_inference_steps=20,
        strength=upscaler_strength,
        control_guidance_start=float(control_guidance_start),
        control_guidance_end=float(control_guidance_end),
        controlnet_conditioning_scale=float(controlnet_conditioning_scale)
    )
    end_time = time.time()
    end_time_struct = time.localtime(end_time)
    end_time_formatted = time.strftime("%H:%M:%S", end_time_struct)
    print(f"Inference ended at {end_time_formatted}, taking {end_time-start_time}s")

    # Save image + metadata
    user_history.save_image(
        label=prompt,
        image=out_image["images"][0],
        profile=profile,
        metadata={
            "prompt": prompt,
            "negative_prompt": negative_prompt,
            "guidance_scale": guidance_scale,
            "controlnet_conditioning_scale": controlnet_conditioning_scale,
            "control_guidance_start": control_guidance_start,
            "control_guidance_end": control_guidance_end,
            "upscaler_strength": upscaler_strength,
            "seed": seed,
            "sampler": sampler,
        },
    )

    return out_image["images"][0], gr.update(visible=True), gr.update(visible=True), my_seed



# Gradio UIの構築

with gr.Blocks() as app:

    # アプリの紹介や説明。テキストやリンクを追加
    gr.Markdown(
        '''
        <div style="text-align: center;">
            <h1>Illusion Diffusion HQ 🌀</h1>
            <p style="font-size:16px;">Stable Diffusion で、驚くほど高品質なイリュージョン・アート作品を生成</p>
            <p>プロンプトとパターンが与えられれば QR コードで調整されたコントロール・ネットを使用して、驚くほど美しいイリュージョンを作成します。</p>
            <p><small>このプロジェクトは、<a href="https://huggingface.co/monster-labs/control_v1p_sd15_qrcode_monster">Monster Labs QR コントロール・ネット</a> を使用して機能します。Illusion Diffusion が安全性チェッカーとともに復活しました!<a href="https://twitter.com/angrypenguinPNG">作者</a> や大きな貢献をしてくれた <a href="https://twitter.com/multimodalart">multimodalart</a> を Twitter でフォローしてください。ワークフローを発見してくれた <a href="https://twitter.com/MrUgleh">MrUgleh</a> に感謝します :)　作者をサポートしたい場合は、<a href="https://deforum.studio">deforum.studio</a> の使用を検討してください。</small></p>
        </div>
        '''
    )

    # 状態の管理
    state_img_input = gr.State()
    state_img_output = gr.State()

    # アプリのレイアウトを設定
    with gr.Row():
        with gr.Column():
            control_image = gr.Image(label="イリュージョンのインプット（画像のレイアウトやストラクチャがわかるモノクロ画像）", type="pil", elem_id="control_image")
            controlnet_conditioning_scale = gr.Slider(minimum=0.0, maximum=5.0, step=0.01, value=0.8, label="イリュージョンの強さ", elem_id="illusion_strength", info="ControlNet 条件付けスケール")
            gr.Examples(examples=["_checkers.png", "_checkers_mid.jpg", "_pattern.png", "_ultra_checkers.png", "_spiral.jpeg", "_funky.jpeg" ], inputs=control_image)
            prompt = gr.Textbox(label="プロンプト（スタイル情報や登場させたいモチーフ）", elem_id="prompt", info="生成したいものを入力してください", placeholder="賑やかな通りと遠くに城がある中世の村の風景")
            negative_prompt = gr.Textbox(label="ネガティブ・プロンプト", info="生成したくないものを入力してください", value="low quality", elem_id="negative_prompt")
            with gr.Accordion(label="高度なオプション", open=False):
                guidance_scale = gr.Slider(minimum=0.0, maximum=50.0, step=0.25, value=7.5, label="ガイダンス・スケール")
                sampler = gr.Dropdown(choices=list(SAMPLER_MAP.keys()), value="Euler")
                control_start = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=0, label="ControlNetの開始")
                control_end = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1, label="ControlNetの終了")
                strength = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1, label="アップスケーラーの強度")
                seed = gr.Slider(minimum=-1, maximum=9999999999, step=1, value=-1, label="シード", info="-1 はランダム・シードです")
                used_seed = gr.Number(label="最後に使用したシード",interactive=False)
            run_btn = gr.Button("実行")
        with gr.Column():
            result_image = gr.Image(label="イリュージョン・ディフュージョンのアウトプット", interactive=False, elem_id="output")
            with gr.Group(elem_id="share-btn-container", visible=False) as share_group:
                community_icon = gr.HTML(community_icon_html)
                loading_icon = gr.HTML(loading_icon_html)
                share_button = gr.Button("コミュニティにシェア", elem_id="share-btn")

    # テキストボックスに入力されたプロンプトが送信されたときに実行されるイベントを設定
    prompt.submit(
        check_inputs,
        inputs=[prompt, control_image],
        queue=False
    ).success(
        inference,
        inputs=[control_image, prompt, negative_prompt, guidance_scale, controlnet_conditioning_scale, control_start, control_end, strength, seed, sampler],
        outputs=[result_image, result_image, share_group, used_seed])

    # 「Run」ボタンがクリックされたときに実行されるイベントを設定
    run_btn.click(
        check_inputs,
        inputs=[prompt, control_image],
        queue=False
    ).success(
        inference,
        inputs=[control_image, prompt, negative_prompt, guidance_scale, controlnet_conditioning_scale, control_start, control_end, strength, seed, sampler],
        outputs=[result_image, result_image, share_group, used_seed])

    # 共有ボタンがクリックされたときに実行されるイベントを設定
    share_button.click(None, [], [], js=share_js)

# アプリケーションの起動
with gr.Blocks(css=css) as app_with_history:
    with gr.Tab("Demo"):
        app.render()
    with gr.Tab("Past generations"):
        user_history.render()

# app_with_history.queue(max_size=20,api_open=True )
# if __name__ == "__main__":
#     app_with_history.launch(max_threads=400)

if __name__ == "__main__":
    app_with_history.queue()
    app_with_history.launch()