app.py

prod = False
port = 8080
show_options = False
if prod:
    port = 8081
    # show_options = False

import os
import random
import time
import gradio as gr
import numpy as np
import spaces
import imageio
from huggingface_hub import HfApi
import gc
import torch
import cv2
from PIL import Image
from diffusers import (
    ControlNetModel,
    DPMSolverMultistepScheduler,
    StableDiffusionControlNetPipeline,
    # StableDiffusionInpaintPipeline,
    # AutoencoderKL,
)
from controlnet_aux_local import NormalBaeDetector

MAX_SEED = np.iinfo(np.int32).max
API_KEY = os.environ.get("API_KEY", None)
# os.environ['HF_HOME'] = '/data/.huggingface'

print("CUDA version:", torch.version.cuda)
print("loading everything")
compiled = False
api = HfApi()

class Preprocessor:
    MODEL_ID = "lllyasviel/Annotators"

    def __init__(self):
        self.model = None
        self.name = ""

    def load(self, name: str) -> None:
        if name == self.name:
            return
        elif name == "NormalBae":
            print("Loading NormalBae")
            self.model = NormalBaeDetector.from_pretrained(self.MODEL_ID).to("cuda")
            torch.cuda.empty_cache()
            self.name = name
        else:
            raise ValueError
        return

    def __call__(self, image: Image.Image, **kwargs) -> Image.Image:
        return self.model(image, **kwargs)

if gr.NO_RELOAD:
    # Controlnet Normal
    model_id = "lllyasviel/control_v11p_sd15_normalbae"
    print("initializing controlnet")
    controlnet = ControlNetModel.from_pretrained(
        model_id,
        torch_dtype=torch.float16,
        attn_implementation="flash_attention_2",
    ).to("cuda")

    # Scheduler
    scheduler = DPMSolverMultistepScheduler.from_pretrained(
        # "runwayml/stable-diffusion-v1-5",
        # "stable-diffusion-v1-5/stable-diffusion-v1-5",
        "ashllay/stable-diffusion-v1-5-archive",
        solver_order=2,
        subfolder="scheduler",
        use_karras_sigmas=True,
        final_sigmas_type="sigma_min",
        algorithm_type="sde-dpmsolver++",
        prediction_type="epsilon",
        thresholding=False,
        denoise_final=True,
        device_map="cuda",
        torch_dtype=torch.float16,
    )

    # Stable Diffusion Pipeline URL
    # base_model_url = "https://huggingface.co/broyang/hentaidigitalart_v20/blob/main/realcartoon3d_v15.safetensors"
    base_model_url = "https://huggingface.co/Lykon/AbsoluteReality/blob/main/AbsoluteReality_1.8.1_pruned.safetensors"
    # vae_url = "https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors"

    # print('loading vae')
    # vae = AutoencoderKL.from_single_file(vae_url, torch_dtype=torch.float16).to("cuda")
    # vae.to(memory_format=torch.channels_last) 

    print('loading pipe')
    pipe = StableDiffusionControlNetPipeline.from_single_file(
        base_model_url,
        safety_checker=None,
        controlnet=controlnet,
        scheduler=scheduler,
        # vae=vae,
        torch_dtype=torch.float16,
    ).to("cuda")
    
    # print('loading inpainting pipe')
    # inpaint_pipe = StableDiffusionInpaintPipeline.from_pretrained(
    #     "runwayml/stable-diffusion-inpainting",
    #     torch_dtype=torch.float16,
    # ).to("cuda")

    print("loading preprocessor")
    preprocessor = Preprocessor()
    preprocessor.load("NormalBae")
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="EasyNegativeV2.safetensors", token="EasyNegativeV2",)
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="badhandv4.pt", token="badhandv4")
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="fcNeg-neg.pt", token="fcNeg-neg")
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="HDA_Ahegao.pt", token="HDA_Ahegao")
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="HDA_Bondage.pt", token="HDA_Bondage")
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="HDA_pet_play.pt", token="HDA_pet_play")
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="HDA_unconventional maid.pt", token="HDA_unconventional_maid")
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="HDA_NakedHoodie.pt", token="HDA_NakedHoodie")
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="HDA_NunDress.pt", token="HDA_NunDress")
    pipe.load_textual_inversion("broyang/hentaidigitalart_v20", weight_name="HDA_Shibari.pt", token="HDA_Shibari")
    pipe.to("cuda")

    print("---------------Loaded controlnet pipeline---------------") 
    torch.cuda.empty_cache()
    gc.collect()
    print(f"CUDA memory allocated: {torch.cuda.max_memory_allocated(device='cuda') / 1e9:.2f} GB")
    print("Model Compiled!")
    
def generate_furniture_mask(image, furniture_type):
    image_np = np.array(image)
    height, width = image_np.shape[:2]
        
    mask = np.zeros((height, width), dtype=np.uint8)
        
    if furniture_type == "sofa":
        cv2.rectangle(mask, (width//4, int(height*0.6)), (width*3//4, height), 255, -1)
    elif furniture_type == "table":
        cv2.rectangle(mask, (width//3, height//3), (width*2//3, height*2//3), 255, -1)
    elif furniture_type == "chair":
        cv2.circle(mask, (width*3//5, height*2//3), height//6, 255, -1)
        
    return Image.fromarray(mask)

def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    return seed

def get_additional_prompt():
    prompt = "hyperrealistic photography,extremely detailed,(intricate details),unity 8k wallpaper,ultra detailed"
    top = ["tank top", "blouse", "button up shirt", "sweater", "corset top"]
    bottom = ["short skirt", "athletic shorts", "jean shorts", "pleated skirt", "short skirt", "leggings", "high-waisted shorts"]
    accessory = ["knee-high boots", "gloves", "Thigh-high stockings", "Garter belt", "choker", "necklace", "headband", "headphones"]
    return f"{prompt}, {random.choice(top)}, {random.choice(bottom)}, {random.choice(accessory)}, score_9"
    # outfit = ["schoolgirl outfit", "playboy outfit", "red dress", "gala dress", "cheerleader outfit", "nurse outfit", "Kimono"]

def get_prompt(prompt, additional_prompt):
    interior = "design-style interior designed (interior space),tungsten white balance,captured with a DSLR camera using f/10 aperture, 1/60 sec shutter speed, ISO 400, 20mm focal length"
    default = "hyperrealistic photography,extremely detailed,(intricate details),unity 8k wallpaper,ultra detailed"
    default2 = f"professional 3d model {prompt},octane render,highly detailed,volumetric,dramatic lighting,hyperrealistic photography,extremely detailed,(intricate details),unity 8k wallpaper,ultra detailed"
    randomize = get_additional_prompt()
    # nude = "NSFW,((nude)),medium bare breasts,hyperrealistic photography,extremely detailed,(intricate details),unity 8k wallpaper,ultra detailed"
    # bodypaint = "((fully naked with no clothes)),nude naked seethroughxray,invisiblebodypaint,rating_newd,NSFW"
    lab_girl = "hyperrealistic photography, extremely detailed, shy assistant wearing minidress boots and gloves, laboratory background, score_9, 1girl"
    pet_play = "hyperrealistic photography, extremely detailed, playful, blush, glasses, collar, score_9, HDA_pet_play"
    bondage = "hyperrealistic photography, extremely detailed, submissive, glasses, score_9, HDA_Bondage"
    # ahegao = "((invisible clothing)), hyperrealistic photography,exposed vagina,sexy,nsfw,HDA_Ahegao"
    ahegao2 = "(invisiblebodypaint),rating_newd,HDA_Ahegao"
    athleisure = "hyperrealistic photography, extremely detailed, 1girl athlete, exhausted embarrassed sweaty,outdoors, ((athleisure clothing)), score_9"
    atompunk = "((atompunk world)), hyperrealistic photography, extremely detailed, short hair, bodysuit, glasses, neon cyberpunk background, score_9"
    maid = "hyperrealistic photography, extremely detailed, shy, blushing, score_9, pastel background, HDA_unconventional_maid"
    nundress = "hyperrealistic photography, extremely detailed, shy, blushing, fantasy background, score_9, HDA_NunDress"
    naked_hoodie = "hyperrealistic photography, extremely detailed, medium hair, cityscape, (neon lights), score_9, HDA_NakedHoodie"
    abg = "(1girl, asian body covered in words, words on body, tattoos of (words) on body),(masterpiece, best quality),medium breasts,(intricate details),unity 8k wallpaper,ultra detailed,(pastel colors),beautiful and aesthetic,see-through (clothes),detailed,solo"
    # shibari = "extremely detailed, hyperrealistic photography, earrings, blushing, lace choker, tattoo, medium hair, score_9, HDA_Shibari"
    shibari2 = "octane render, highly detailed, volumetric, HDA_Shibari"
    
    if prompt == "":
        girls = [randomize, pet_play, bondage, lab_girl, athleisure, atompunk, maid, nundress, naked_hoodie, abg, shibari2, ahegao2]
        prompts_nsfw = [abg, shibari2, ahegao2]
        prompt = f"{random.choice(girls)}"
        prompt = f"boho chic"
        # print(f"-------------{preset}-------------")
    else:
        prompt = f"Photo from Pinterest of {prompt} {interior}"
        # prompt = default2
    return f"{prompt} f{additional_prompt}"

style_list = [
    {
        "name": "None",
        "prompt": ""
    },
    {
        "name": "Minimalistic",
        "prompt": "Minimalist interior design,clean lines,neutral colors,uncluttered space,functional furniture,lots of natural light"
    },
    {
        "name": "Boho",
        "prompt": "Bohemian chic interior,eclectic mix of patterns and textures,vintage furniture,plants,woven textiles,warm earthy colors"
    },
    {
        "name": "Farmhouse",
        "prompt": "Modern farmhouse interior,rustic wood elements,shiplap walls,neutral color palette,industrial accents,cozy textiles"
    },
    {
        "name": "Saudi Prince",
        "prompt": "Opulent gold interior,luxurious ornate furniture,crystal chandeliers,rich fabrics,marble floors,intricate Arabic patterns"
    },
    {
        "name": "Neoclassical",
        "prompt": "Neoclassical interior design,elegant columns,ornate moldings,symmetrical layout,refined furniture,muted color palette"
    },
    {
        "name": "Eclectic",
        "prompt": "Eclectic interior design,mix of styles and eras,bold color combinations,diverse furniture pieces,unique art objects"
    },
    {
        "name": "Parisian",
        "prompt": "Parisian apartment interior,all-white color scheme,ornate moldings,herringbone wood floors,elegant furniture,large windows"
    },
    {
        "name": "Hollywood",
        "prompt": "Hollywood Regency interior,glamorous and luxurious,bold colors,mirrored surfaces,velvet upholstery,gold accents"
    },
    {
        "name": "Scandinavian",
        "prompt": "Scandinavian interior design,light wood tones,white walls,minimalist furniture,cozy textiles,hygge atmosphere"
    },
    {
        "name": "Beach",
        "prompt": "Coastal beach house interior,light blue and white color scheme,weathered wood,nautical accents,sheer curtains,ocean view"
    },
    {
        "name": "Japanese",
        "prompt": "Traditional Japanese interior,tatami mats,shoji screens,low furniture,zen garden view,minimalist decor,natural materials"
    },
    { 
        "name": "Midcentury Modern",
        "prompt": "Mid-century modern interior,1950s-60s style furniture,organic shapes,warm wood tones,bold accent colors,large windows"
    },
    {
        "name": "Retro Futurism",
        "prompt": "Neon (atompunk world) retro cyberpunk background",
    },
    {
        "name": "Texan",
        "prompt": "Western cowboy interior,rustic wood beams,leather furniture,cowhide rugs,antler chandeliers,southwestern patterns"
    },
    {
        "name": "Matrix",
        "prompt": "Futuristic cyberpunk interior,neon accent lighting,holographic plants,sleek black surfaces,advanced gaming setup,transparent screens,Blade Runner inspired decor,high-tech minimalist furniture"
    }
] 

styles = {k["name"]: (k["prompt"]) for k in style_list}
STYLE_NAMES = list(styles.keys())

def apply_style(style_name):
    if style_name in styles:
        p = styles.get(style_name, "none")
    return p

    
css = """
h1, h2, h3 {
    text-align: center;
    display: block;
}
footer {
    visibility: hidden;
}
.gradio-container {
    max-width: 1100px !important;
}
.gr-image {
    display: flex;
    justify-content: center; 
    align-items: center;
    width: 100%;
    height: 512px;
    overflow: hidden;
}
.gr-image img {
    width: 100%;
    height: 100%; 
    object-fit: cover;
    object-position: center;
}
"""
with gr.Blocks(theme="bethecloud/storj_theme", css=css) as demo:
    #############################################################################
    with gr.Row():
        with gr.Accordion("Advanced options", open=show_options, visible=show_options):
            num_images = gr.Slider(
                label="Images", minimum=1, maximum=4, value=1, step=1
            )
            image_resolution = gr.Slider(
                label="Image resolution",
                minimum=256,
                maximum=1024,
                value=512,
                step=256,
            )
            preprocess_resolution = gr.Slider(
                label="Preprocess resolution",
                minimum=128,
                maximum=1024,
                value=512,
                step=1,
            )
            num_steps = gr.Slider(
                label="Number of steps", minimum=1, maximum=100, value=15, step=1
            )  # 20/4.5 or 12 without lora, 4 with lora
            guidance_scale = gr.Slider(
                label="Guidance scale", minimum=0.1, maximum=30.0, value=5.5, step=0.1
            )  # 5 without lora, 2 with lora
            seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
            a_prompt = gr.Textbox(
                label="Additional prompt",
                value = "design-style interior designed (interior space), tungsten white balance, captured with a DSLR camera using f/10 aperture, 1/60 sec shutter speed, ISO 400, 20mm focal length"
            )
            n_prompt = gr.Textbox(
                label="Negative prompt",
                value="EasyNegativeV2, fcNeg, (badhandv4:1.4), (worst quality, low quality, bad quality, normal quality:2.0), (bad hands, missing fingers, extra fingers:2.0)",
            )
    #############################################################################
    # input text
    with gr.Column():
        prompt = gr.Textbox(
            label="Custom Design",
            placeholder="Enter a description (optional)",
        )
    # design options
    with gr.Row(visible=True):
        style_selection = gr.Radio(
            show_label=True,
            container=True,
            interactive=True,
            choices=STYLE_NAMES,
            value="None",
            label="Design Styles",
        )
    # input image
    with gr.Row(equal_height=True):
        with gr.Column(scale=1, min_width=300):
            image = gr.Image(
                label="Input",
                sources=["upload"],
                show_label=True,
                mirror_webcam=True,
                type="pil",
            )
            # run button
            with gr.Column():
                run_button = gr.Button(value="Use this one", size="lg", visible=False)
        # output image
        with gr.Column(scale=1, min_width=300):
            result = gr.Image(  
                label="Output",
                interactive=False,
                type="pil",
                show_share_button= False,
            )
            # Use this image button
            with gr.Column():
                use_ai_button = gr.Button(value="Use this one", size="lg", visible=False)
    config = [
        image,
        style_selection,
        prompt,
        a_prompt,
        n_prompt,
        num_images,
        image_resolution,
        preprocess_resolution,
        num_steps,
        guidance_scale,
        seed,
    ]
    
    with gr.Row():
        helper_text = gr.Markdown("## Tap and hold (on mobile) to save the image.", visible=True)
    
    # image processing
    @gr.on(triggers=[image.upload, prompt.submit, run_button.click], inputs=config, outputs=result, show_progress="minimal")
    def auto_process_image(image, style_selection, prompt, a_prompt, n_prompt, num_images, image_resolution, preprocess_resolution, num_steps, guidance_scale, seed, progress=gr.Progress(track_tqdm=True)):
        return process_image(image, style_selection, prompt, a_prompt, n_prompt, num_images, image_resolution, preprocess_resolution, num_steps, guidance_scale, seed)
    
    # AI image processing
    @gr.on(triggers=[use_ai_button.click], inputs=[result] + config, outputs=[image, result], show_progress="minimal")
    def submit(previous_result, image, style_selection, prompt, a_prompt, n_prompt, num_images, image_resolution, preprocess_resolution, num_steps, guidance_scale, seed, progress=gr.Progress(track_tqdm=True)):
        # First, yield the previous result to update the input image immediately
        yield previous_result, gr.update()
        # Then, process the new input image
        new_result = process_image(previous_result, style_selection, prompt, a_prompt, n_prompt, num_images, image_resolution, preprocess_resolution, num_steps, guidance_scale, seed)
        # Finally, yield the new result
        yield previous_result, new_result

    # Turn off buttons when processing
    @gr.on(triggers=[image.upload, use_ai_button.click, run_button.click], inputs=None, outputs=[run_button, use_ai_button], show_progress="hidden")
    def turn_buttons_off():
        return gr.update(visible=False), gr.update(visible=False)
    
    # Turn on buttons when processing is complete
    @gr.on(triggers=[result.change], inputs=None, outputs=[use_ai_button, run_button], show_progress="hidden")
    def turn_buttons_on():
        return gr.update(visible=True), gr.update(visible=True)

@spaces.GPU(duration=12)
@torch.inference_mode()
def process_image(
    image,
    style_selection,
    prompt,
    a_prompt,
    n_prompt,
    num_images,
    image_resolution,
    preprocess_resolution,
    num_steps,
    guidance_scale,
    seed,
):
    seed = random.randint(0, MAX_SEED)
    generator = torch.cuda.manual_seed(seed)
    
    preprocessor.load("NormalBae")
    control_image = preprocessor(
        image=image,
        image_resolution=image_resolution,
        detect_resolution=preprocess_resolution,
    )
    
    if style_selection is not None and style_selection != "None":
        prompt = f"Photo from Pinterest of {apply_style(style_selection)} {prompt},{a_prompt}"
    else:
        prompt = str(get_prompt(prompt, a_prompt))
    negative_prompt = str(n_prompt)
    print(prompt)
    
    # Generate the initial room image
    initial_result = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        guidance_scale=guidance_scale,
        num_images_per_prompt=1,
        num_inference_steps=num_steps,
        generator=generator,
        image=control_image,
    ).images[0]
    
    # # Randomly choose whether to add furniture and which type
    # furniture_types = ["sofa", "table", "chair", "dresser", "bookshelf", "desk", "coffee table"]
    # furniture_type = random.choice(furniture_types)
    
    
    # furniture_mask = generate_furniture_mask(initial_result, furniture_type)
    # furniture_prompt = f"{prompt}, with a {furniture_type} in the style of {style_selection}"
    # print(furniture_prompt)
    
    # # Use the inpainting pipeline to add furniture
    # final_result = inpaint_pipe(
    #     prompt=furniture_prompt,
    #     image=initial_result,
    #     mask_image=furniture_mask,
    #     negative_prompt=negative_prompt,
    #     num_inference_steps=num_steps,
    #     guidance_scale=guidance_scale,
    #     generator=generator,
    # ).images[0]
    
    # Save and upload results
    timestamp = int(time.time())
    img_path = f"{timestamp}.jpg"
    results_path = f"{timestamp}_out.jpg"
    imageio.imsave(img_path, image)
    imageio.imsave(results_path, initial_result)
    api.upload_file(
        path_or_fileobj=img_path,
        path_in_repo=img_path,
        repo_id="broyang/interior-ai-outputs",
        repo_type="dataset",
        token=API_KEY,
        run_as_future=True,
    )
    api.upload_file(
        path_or_fileobj=results_path,
        path_in_repo=results_path,
        repo_id="broyang/interior-ai-outputs",
        repo_type="dataset",
        token=API_KEY,
        run_as_future=True,
    )
    return initial_result

if prod:
    demo.queue(max_size=20).launch(server_name="localhost", server_port=port)
else:
    demo.queue(api_open=False).launch(show_api=False)