app.py

import gradio as gr
import numpy as np
from PIL import Image
import torch
from diffusers import ControlNetModel, UniPCMultistepScheduler
from hico_pipeline import StableDiffusionControlNetMultiLayoutPipeline

device = "cuda" if torch.cuda.is_available() else "cpu"

# Initialize model
controlnet = ControlNetModel.from_pretrained("qihoo360/HiCo_T2I", torch_dtype=torch.float16)
pipe = StableDiffusionControlNetMultiLayoutPipeline.from_pretrained(
    "krnl/realisticVisionV51_v51VAE", controlnet=[controlnet], torch_dtype=torch.float16
)
pipe = pipe.to(device)
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)

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

# Store objects
object_classes_list = ["A photograph of a young woman wrapped in a towel wearing a pair of sunglasses", "a towel", "a young woman wrapped in a towel wearing a pair of sunglasses", "a pair of sunglasses"]
object_bboxes_list = ["0,0,512,512", "17,77,144,155", "16,28,157,155", "82,44,129,63"]

# Function to add or update the prompt in the list
def submit_prompt(prompt):
    if object_classes_list:
        object_classes_list[0] = prompt  # Overwrite the first element if it exists
    else:
        object_classes_list.insert(0, prompt)  # Add to the beginning if the list is empty

    if not object_bboxes_list:
        object_bboxes_list.insert(0, "0,0,512,512")  # Add the default bounding box if the list is empty

    combined_list = [[cls, bbox] for cls, bbox in zip(object_classes_list, object_bboxes_list)]
    return combined_list, gr.update(interactive=False)  # Make the prompt input non-editable

# Function to add a new object with validation
def add_object(object_class, bbox):
    try:
        # Split and convert bbox string into integers
        x1, y1, x2, y2 = map(int, bbox.split(","))
        
        # Validate the coordinates
        if x2 < x1 or y2 < y1:
            return "Error: x2 cannot be less than x1 and y2 cannot be less than y1.", []
        if x1 < 0 or y1 < 0 or x2 > 512 or y2 > 512:
            return "Error: Coordinates must be between 0 and 512.", []

        # If validation passes, add to the lists
        object_classes_list.append(object_class)
        object_bboxes_list.append(bbox)
        combined_list = [[cls, bbox] for cls, bbox in zip(object_classes_list, object_bboxes_list)]
        return combined_list

    except ValueError:
        return "Error: Invalid input format. Use x1,y1,x2,y2.", []

# Function to generate images based on added objects
def generate_image(prompt, guidance_scale, num_inference_steps, randomize_seed, seed):
    img_width, img_height = 512, 512
    r_image = np.zeros((img_height, img_width, 3), dtype=np.uint8)
    list_cond_image = []

    for bbox in object_bboxes_list:
        x1, y1, x2, y2 = map(int, bbox.split(","))
        cond_image = np.zeros_like(r_image, dtype=np.uint8)
        cond_image[y1:y2, x1:x2] = 255
        list_cond_image.append(Image.fromarray(cond_image).convert('RGB'))

    if randomize_seed or seed is None:
        seed = np.random.randint(0, MAX_SEED)

    generator = torch.manual_seed(seed)

    image = pipe(
        prompt=prompt,
        layo_prompt=object_classes_list,
        guess_mode=False,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        image=list_cond_image,
        fuse_type="avg",
        width=512,
        height=512
    ).images[0]
    print(type(image),'image')
    return image, seed

# Function to clear all arrays and reset the UI
def clear_arrays():
    object_classes_list.clear()
    object_bboxes_list.clear()
    return [], gr.update(value="", interactive=True)  # Clear the objects and reset the prompt

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("# HiCo_T2I 512px")
    gr.Markdown(" You can directly click **Generate Image** or customize it by first entering the global caption, followed by subcaptions and their corresponding coordinates.")


    # Put prompt and submit button in the same row
    with gr.Group():
        with gr.Row():
            # Prompt input and submit button
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt here",
                container=False,
                
            )
            submit_button = gr.Button("Submit Prompt", scale=0)

    # Always visible DataFrame
    objects_display = gr.Dataframe(
        headers=["Caption", "Bounding Box"],
        value=[[cls, bbox] for cls, bbox in zip(object_classes_list, object_bboxes_list)]
    )

    with gr.Row():
        object_class_input = gr.Textbox(label="Sub-caption", placeholder="Enter Sub-caption (e.g., apple)")
        bbox_input = gr.Textbox(label="Bounding Box (x1,y1,x2,y2 and >=0 and <=512)", placeholder="Enter bounding box coordinates")

    add_button = gr.Button("Add")

    # Advanced settings in a collapsible accordion
    with gr.Accordion("Advanced Settings", open=False):
        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():
            guidance_scale = gr.Slider(
                label="Guidance scale",
                minimum=0.0,
                maximum=10.0,
                step=0.1,
                value=7.5
            )
            num_inference_steps = gr.Slider(
                label="Number of inference steps",
                minimum=1,
                maximum=50,
                step=1,
                value=50
            )

    generate_button = gr.Button("Generate Image")
    result = gr.Image(label="Generated Image")

    # Refresh button to clear arrays and reset inputs (moved below the result)
    refresh_button = gr.Button("Refresh")

    # Submit the prompt and update the display
    submit_button.click(
        fn=submit_prompt,
        inputs=prompt,
        outputs=[objects_display, prompt]
    )

    # Add object and update display
    add_button.click(
        fn=add_object,
        inputs=[object_class_input, bbox_input],
        outputs=[objects_display]
    )

    # Generate image based on added objects
    generate_button.click(
        fn=generate_image,
        inputs=[prompt, guidance_scale, num_inference_steps, randomize_seed, seed],
        outputs=[result, seed]
    )

    # Refresh button to clear arrays and reset inputs
    refresh_button.click(
        fn=clear_arrays,
        inputs=None,
        outputs=[objects_display, prompt]
    )

if __name__ == "__main__":
    demo.launch()