app.py

import gradio as gr
import torch
from diffusers import StableDiffusionPipeline, StableDiffusionControlNetPipeline, ControlNetModel, EulerDiscreteScheduler
from jinja2 import Template
import numpy as np
import cv2
from PIL import Image
import time


# Load models from Hugging Face
sd_model_id = "bhoomikagp/sd2-interior-model-version2" ## test
# sd_model_id = "bhoomikagp/sd3-interior-model" ## SD3 model issue loading
controlnet_model_id = "lllyasviel/sd-controlnet-mlsd"
# Load Stable Diffusion pipeline CUDA
scheduler = EulerDiscreteScheduler.from_pretrained(sd_model_id, subfolder="scheduler")
#sd_pipeline = StableDiffusionPipeline.from_pretrained(sd_model_id, torch_dtype=torch.float16,scheduler=scheduler).to("cuda")
# Load ControlNet and Stable Diffusion ControlNet pipeline
"""
controlnet = ControlNetModel.from_pretrained(controlnet_model_id, torch_dtype=torch.float16).to("cuda")
controlnet_pipeline = StableDiffusionControlNetPipeline.from_pretrained(
    sd_model_id,
    controlnet=controlnet,
    scheduler=scheduler,
    torch_dtype=torch.float16
).to("cuda")
"""
#sd_model_id = "stabilityai/stable-diffusion-2-1"
#scheduler = DPMSolverMultistepScheduler.from_pretrained(sd_model_id, subfolder="scheduler")

# Check if CUDA is available
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")

# Initialize the pipeline with appropriate device and dtype
torch_dtype = torch.float16 if device == "cuda" else torch.float32
sd_pipeline = StableDiffusionPipeline.from_pretrained(
    sd_model_id,
    torch_dtype=torch_dtype,
    scheduler=scheduler
).to(device)

# choices lists
option_choices = ["living_room", "bedroom", "kitchen"]
adj_1_choices = [
    "spacious", "cozy", "minimalist", "elegant", "modern", "rustic",
    "luxurious", "inviting", "airy", "sophisticated", "bright", "warm",
    "serene", "chic", "contemporary"
]
architecture_style_choices = [
    "modern", "contemporary", "traditional", "industrial", "scandinavian",
    "mid-century", "colonial", "art deco", "neo-classical", "mediterranean",
    "gothic", "baroque", "japanese", "brutalist", "tropical"
]
aesthetic_choices = [
    "bohemian", "vintage", "minimalist", "luxurious", "eclectic",
    "mid-century", "art deco", "modern farmhouse", "industrial chic",
    "shabby chic", "rustic elegance", "coastal", "urban", "transitional",
    "contemporary classic"
]
primary_color_choices = [
    "blue", "green", "beige", "grey", "white", "black",
    "cream", "brown", "taupe", "burgundy", "mustard",
    "terracotta", "olive", "peach", "navy"
]
wood_finish_choices = [
    "dark oak", "walnut", "mahogany", "teak", "maple",
    "cherry", "pine", "birch", "ash", "rosewood",
    "ebony", "cedar", "hickory", "elm", "red oak"
]
wall_color_choices = [
    "cream", "off-white", "charcoal", "sage green", "navy blue",
    "taupe", "light grey", "soft pink", "mustard yellow", "deep teal",
    "warm beige", "pearl white", "slate blue", "coral", "mint green"
]
tiles_choices = [
    "marble", "ceramic", "porcelain", "slate", "wooden-look",
    "mosaic", "granite", "terracotta", "cement", "quartz",
    "limestone", "onyx", "travertine", "glass", "encaustic"
]


# Templates for each room type
templates = {
    "living_room": """
    High quality, High resolution, Interior, Architecture, Revit, Autocad, Realistic 3D Render, vray, lumion, raytracing,
    of a {{ adj_1 }} living room, in {{ architecture_style }} architecture, with {{ aesthetic }} style,
    painted in {{ wall_color }} with {{ primary_color }} accents, {{ wood_finish }} wood finishes, and {{ tiles }} flooring.
    """,
    "bedroom": """
    High quality, High resolution, Interior, Architecture, Revit, Autocad, Realistic 3D Render, vray, lumion, raytracing,
    of a {{ adj_1 }} bedroom, in {{ architecture_style }} architecture, with {{ aesthetic }} style,
    painted in {{ wall_color }} with {{ primary_color }} accents, {{ wood_finish }} wood finishes, and {{ tiles }} flooring.
    """,
    "kitchen": """
    High quality, High resolution, Interior, Architecture, Revit, Autocad, Realistic 3D Render, vray, lumion, raytracing,
    of a {{ adj_1 }} kitchen, in {{ architecture_style }} architecture, with {{ aesthetic }} style,
    painted in {{ wall_color }} with {{ primary_color }} accents, {{ wood_finish }} cabinetry, and {{ tiles }} flooring.
    """
}

def generate_prompt(option, adj_1, architecture_style, aesthetic, primary_color, wood_finish, wall_color, tiles):
    # Validate inputs
    if adj_1 not in adj_1_choices:
        raise ValueError(f"Invalid adjective: '{adj_1}'. Accepted options are {adj_1_choices}")
    if architecture_style not in architecture_style_choices:
        raise ValueError(f"Invalid architecture style: '{architecture_style}'. Accepted options are {architecture_style_choices}")
    if aesthetic not in aesthetic_choices:
        raise ValueError(f"Invalid aesthetic style: '{aesthetic}'. Accepted options are {aesthetic_choices}")
    if primary_color not in primary_color_choices:
        raise ValueError(f"Invalid primary color: '{primary_color}'. Accepted options are {primary_color_choices}")
    if wood_finish not in wood_finish_choices:
        raise ValueError(f"Invalid wood finish: '{wood_finish}'. Accepted options are {wood_finish_choices}")
    if wall_color not in wall_color_choices:
        raise ValueError(f"Invalid wall color: '{wall_color}'. Accepted options are {wall_color_choices}")
    if tiles not in tiles_choices:
        raise ValueError(f"Invalid tiles: '{tiles}'. Accepted options are {tiles_choices}")

    # Select the template based on the room type option
    template_str = templates.get(option.lower())
    if not template_str:
        raise ValueError(f"Invalid option: '{option}'. Available options are {option_choices}")

    # Render the template
    template = Template(template_str)
    return template.render(
        adj_1=adj_1,
        architecture_style=architecture_style,
        aesthetic=aesthetic,
        primary_color=primary_color,
        wood_finish=wood_finish,
        wall_color=wall_color,
        tiles=tiles
    )

# Function to generate initial image
def generate_initial_image(n_images=1, option='living_room',  adj_1="spacious", architecture_style="modern", aesthetic="minimalist",
                      primary_color="neutral", wood_finish="oak", wall_color="off-white", tiles="marble"):
    # Generate the prompt from choices
    gen_prompt = generate_prompt(option, adj_1, architecture_style, aesthetic, primary_color, wood_finish, wall_color, tiles)

    images = []  # List to store generated images along with seeds
    for i in range(n_images):
        # Generate a random seed for reproducibility
        seed = torch.randint(0, 2**32, (1,)).item()
        generator = torch.manual_seed(seed)

        # Generate the image with the prompt and Stable Diffusion pipeline
        cfg = 8 # Configuring guidance scale
        steps = 30  # Number of inference steps
        width, height = 640, 512
        # pipeline without controlnet
        img = sd_pipeline(prompt=gen_prompt, guidance_scale=cfg, num_inference_steps=steps, width=width, height=height, generator=generator).images[0]

        # Append the seed and image to the list for tracking
        images.append((seed, img))
    return images

# Function to modify image using ControlNet pipeline
def modify_image_with_controlnet(selected_image, additional_prompt, option='living_room',  adj_1="spacious", architecture_style="modern", aesthetic="minimalist",
                      primary_color="brown", wood_finish="walnut", wall_color="beige", tiles="ceramic"):
    gen_prompt = generate_prompt(option, adj_1, architecture_style, aesthetic, primary_color, wood_finish, wall_color, tiles)
    gen_prompt += '(('+additional_prompt+'))'
    seed = selected_image[0]
    generator = torch.manual_seed(seed)
    cfg = 8  # Configuring guidance scale
    steps = 30  # Number of inference steps
    width, height = 640, 512
    # preprocessig for control net
    img = np.array(selected_image[1])
    gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
    edges = cv2.Canny(gray, 100, 200)
    edges_image = Image.fromarray(edges)
    controlnet_strength = 0.85
    image = controlnet_pipeline(prompt=gen_prompt, image=edges_image, guidance_scale=cfg, width=width, height=height,
                 controlnet_conditioning_scale=controlnet_strength, num_inference_steps=steps, generator=generator).images[0]
    return image

# Gradio app logic
def generate_image(option, adj_1, architecture_style, aesthetic, primary_color, wood_finish, wall_color, tiles, n_images, progress=gr.Progress()):
    generated_images = generate_initial_image(
        n_images=n_images,
        option=option,
        adj_1=adj_1,
        architecture_style=architecture_style,
        aesthetic=aesthetic,
        primary_color=primary_color,
        wood_finish=wood_finish,
        wall_color=wall_color,
        tiles=tiles
    )

    display_images = [img.resize((640, 512)) for _, img in generated_images]  # Resize only for display
    image_identifiers = [f"Image-{i+1}" for i in range(n_images)]

    return display_images, generated_images, image_identifiers


def modify_image(selected_image_id, additional_prompt, option, adj_1, architecture_style, aesthetic, primary_color, wood_finish, wall_color, tiles, images):
    # Parse selected image identifier to get the index
    image_index = int(selected_image_id.split("-")[1]) - 1
    selected_image = images[image_index]  # Retrieve the (seed, original_img) tuple without resizing

    # Modify image using ControlNet
    modified_image = modify_image_with_controlnet(
        selected_image=selected_image,
        additional_prompt=additional_prompt,
        option=option,
        adj_1=adj_1,
        architecture_style=architecture_style,
        aesthetic=aesthetic,
        primary_color=primary_color,
        wood_finish=wood_finish,
        wall_color=wall_color,
        tiles=tiles
    )
    return modified_image

# Interface
with gr.Blocks() as app:
    gr.Markdown("# Interior Design Image Generation and Modification")

    # Image generation options
    with gr.Row():
        option = gr.Radio(label="Room Type", choices=option_choices, value="living_room")
        adj_1 = gr.Radio(label="Primary Adjective", choices=adj_1_choices, value="spacious")
        architecture_style = gr.Radio(label="Architecture Style", choices=architecture_style_choices, value="modern")

    with gr.Row():
        aesthetic = gr.Radio(label="Aesthetic", choices=aesthetic_choices, value="minimalist")
        primary_color = gr.Radio(label="Primary Color", choices=primary_color_choices, value="neutral")
        wood_finish = gr.Radio(label="Wood Finish", choices=wood_finish_choices, value="oak")

    with gr.Row():
        wall_color = gr.Radio(label="Wall Color", choices=wall_color_choices, value="off-white")
        tiles = gr.Radio(label="Tile Type", choices=tiles_choices, value="marble")

    n_images = gr.Slider(label="Number of Images to Generate", minimum=1, maximum=5, step=1, value=1)

    # Button to generate images
    generate_btn = gr.Button("Generate Images")
    output_images = gr.Gallery(label="Generated Images", columns=5)  # Display resized images in gallery layout
    output_generated_images = gr.State()  # Hidden state to store original (unmodified) images data
    output_identifiers = gr.State()  # Hidden state for image identifiers

    # Image selection and modification
    selected_image_id = gr.Radio(label="Select an Image for Modification", choices=[])
    additional_prompt = gr.Textbox(label="Additional Prompt for Modification", placeholder="OPTIONAL")
    modify_btn = gr.Button("Modify Image")
    modified_image = gr.Image(label="Modified Image")

    # Button actions
    generate_btn.click(
        fn=generate_image,
        inputs=[option, adj_1, architecture_style, aesthetic, primary_color, wood_finish, wall_color, tiles, n_images],
        outputs=[output_images, output_generated_images, output_identifiers]
    )

    # Update radio button options after images are generated
    output_images.change(
        fn=lambda identifiers: gr.update(choices=identifiers, visible=True),  # Populate with image identifiers
        inputs=[output_identifiers],
        outputs=[selected_image_id]
    )

    # Trigger modification on button click
    modify_btn.click(
        fn=modify_image,
        inputs=[selected_image_id, additional_prompt, option, adj_1, architecture_style, aesthetic, primary_color, wood_finish, wall_color, tiles, output_generated_images],
        outputs=modified_image
    )

app.launch(debug=True, share=True)