app.py

from diffusers import DDIMScheduler
from pipeline_stable_diffusion_xl_opt import StableDiffusionXLPipeline
from injection_utils import regiter_attention_editor_diffusers
from bounded_attention import BoundedAttention
from pytorch_lightning import seed_everything

import gradio as gr
import torch
import numpy as np
from PIL import Image, ImageDraw

from functools import partial

RESOLUTION = 512
MIN_SIZE = 0.01
WHITE = 255
COLORS = ["red", "blue", "green", "orange", "purple", "turquoise", "olive"]


def inference(
    device,
    model,
    boxes,
    prompts,
    subject_token_indices,
    filter_token_indices,
    num_tokens,
    init_step_size,
    final_step_size,
    num_clusters_per_subject,
    cross_loss_scale,
    self_loss_scale,
    classifier_free_guidance_scale,
    num_iterations,
    loss_threshold,
    num_guidance_steps,
    seed,
):
    seed_everything(seed)
    start_code = torch.randn([len(prompts), 4, 128, 128], device=device)
    editor = BoundedAttention(
        boxes,
        prompts,
        subject_token_indices,
        list(range(70, 82)),
        list(range(70, 82)),
        eos_token_index=num_tokens + 1,
        cross_loss_coef=cross_loss_scale,
        self_loss_coef=self_loss_scale,
        filter_token_indices=filter_token_indices,
        max_guidance_iter=num_guidance_steps,
        max_guidance_iter_per_step=num_iterations,
        start_step_size=init_step_size,
        end_step_size=final_step_size,
        loss_stopping_value=loss_threshold,
        num_clusters_per_box=num_clusters_per_subject,
        debug=False,
    )

    regiter_attention_editor_diffusers(model, editor)
    return model(prompts, latents=start_code, guidance_scale=classifier_free_guidance_scale).images


def generate(
    device,
    model,
    prompt,
    subject_token_indices,
    filter_token_indices,
    num_tokens,
    init_step_size,
    final_step_size,
    num_clusters_per_subject,
    cross_loss_scale,
    self_loss_scale,
    classifier_free_guidance_scale,
    batch_size,
    num_iterations,
    loss_threshold,
    num_guidance_steps,
    seed,
    boxes
):
    subject_token_indices = convert_token_indices(subject_token_indices, nested=True)
    if len(boxes) != len(subject_token_indices):
        raise ValueError("""
            The number of boxes should be equal to the number of subject token indices.
            Number of boxes drawn: {}, number of grounding tokens: {}.
        """.format(len(boxes), len(subject_token_indices)))

    filter_token_indices = convert_token_indices(filter_token_indices) if len(filter_token_indices.strip()) > 0 else None
    num_tokens = int(num_tokens) if len(num_tokens.strip()) > 0 else None
    prompts = [prompt.strip('.').strip(',').strip()] * batch_size

    images = inference(
        device, model, boxes, prompts, subject_token_indices, filter_token_indices, num_tokens, init_step_size,
        final_step_size, num_clusters_per_subject, cross_loss_scale, self_loss_scale, classifier_free_guidance_scale,
        num_iterations, loss_threshold, num_guidance_steps, seed)

    blank_samples = batch_size % 2 if batch_size > 1 else 0
    images = [gr.Image.update(value=x, visible=True) for i, x in enumerate(images)] \
                 + [gr.Image.update(value=None, visible=True) for _ in range(blank_samples)] \
                 + [gr.Image.update(value=None, visible=False) for _ in range(4 - batch_size - blank_samples)]

    return images


def convert_token_indices(token_indices, nested=False):
    if nested:
        return [convert_token_indices(indices, nested=False) for indices in token_indices.split(';')]

    return [int(index.strip()) for index in token_indices.split(',') if len(index.strip()) > 0]


def draw(boxes, mask, layout_image):
    if mask.ndim == 3:
        mask = WHITE - mask[..., 0]

    mask = (mask != 0).astype('uint8') * WHITE
    if mask.sum() > 0:
        x1x2 = np.where(mask.max(0) != 0)[0] / RESOLUTION
        y1y2 = np.where(mask.max(1) != 0)[0] / RESOLUTION
        y1, y2 = y1y2.min(), y1y2.max()
        x1, x2 = x1x2.min(), x1x2.max()

        if (x2 - x1 > MIN_SIZE) and (y2 - y1 > MIN_SIZE):
            boxes.append((x1, y1, x2, y2))
            layout_image = draw_boxes(boxes)

    return [boxes, None, layout_image]


def draw_boxes(boxes):
    if len(boxes) == 0:
        return None

    boxes = np.array(boxes) * RESOLUTION
    image = Image.new('RGB', (RESOLUTION, RESOLUTION), (WHITE, WHITE, WHITE))
    drawing = ImageDraw.Draw(image)
    for i, box in enumerate(boxes.astype(int).tolist()):
        drawing.rectangle(box, outline=COLORS[i % len(COLORS)], width=4)

    return image


def clear(batch_size):
    blank_samples = batch_size % 2 if batch_size > 1 else 0
    out_images = [gr.Image.update(value=None, visible=True) for i in range(batch_size)]
    return [[], None, None] + out_images


def main():
    css = """
    #paper-info a {
        color:#008AD7;
        text-decoration: none;
    }
    #paper-info a:hover {
        cursor: pointer;
        text-decoration: none;
    }

    .tooltip {
        color: #555;
        position: relative;
        display: inline-block;
        cursor: pointer;
    }

    .tooltip .tooltiptext {
        visibility: hidden;
        width: 400px;
        background-color: #555;
        color: #fff;
        text-align: center;
        padding: 5px;
        border-radius: 5px;
        position: absolute;
        z-index: 1; /* Set z-index to 1 */
        left: 10px;
        top: 100%;
        opacity: 0;
        transition: opacity 0.3s;
    }

    .tooltip:hover .tooltiptext {
        visibility: visible;
        opacity: 1;
        z-index: 9999; /* Set a high z-index value when hovering */
    }
    """

    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    model_path = "stabilityai/stable-diffusion-xl-base-1.0"
    scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
    model = StableDiffusionXLPipeline.from_pretrained(model_path, scheduler=scheduler, torch_dtype=torch.float16).to(device)
    model.unet.set_default_attn_processor()
    model.enable_xformers_memory_efficient_attention()
    model.enable_sequential_cpu_offload()

    with gr.Blocks(
            css=css,
            title="Bounded Attention demo",
    ) as demo:
        description = """<p style="text-align: center; font-weight: bold;">
            <span style="font-size: 28px">Bounded Attention</span>
            <br>
            <span style="font-size: 18px" id="paper-info">
                [<a href="https://omer11a.github.io/bounded-attention/" target="_blank">Project Page</a>]
                [<a href=" " target="_blank">Paper</a>]
                [<a href="https://github.com/omer11a/bounded-attention" target="_blank">GitHub</a>]
            </span>
        </p>
        """
        gr.HTML(description)
        with gr.Column():
            prompt = gr.Textbox(
                label="Text prompt",
            )

            subject_token_indices = gr.Textbox(
                label="The token indices of each subject (separate indices for the same subject with commas, and between different subjects with semicolons)",
            )

            filter_token_indices = gr.Textbox(
                label="The token indices to filter, i.e. conjunctions, number, postional relations, etc. (if left empty, this will be automatically inferred)",
            )

            num_tokens = gr.Textbox(
                label="The number of tokens in the prompt (can be left empty, but we recommend filling this, so we can verify your input, as sometimes rare words are split into more than one token)",
            )

            with gr.Row():
                sketch_pad = gr.Sketchpad(label="Sketch Pad", shape=(RESOLUTION, RESOLUTION))
                layout_image = gr.Image(type="pil", label="Bounding Boxes")
                out_images = gr.Image(type="pil", visible=True, label="Generated Image")

            with gr.Row():
                clear_button = gr.Button(value='Clear')
                generate_button = gr.Button(value='Generate')

            with gr.Accordion("Advanced Options", open=False):
                with gr.Column():
                    description = """
                        <div class="tooltip">Batch size &#9432
                        <span class="tooltiptext">The number of images to generate.</span>
                        </div>
                        <div class="tooltip">Initial step size &#9432
                        <span class="tooltiptext">The initial step size of the linear step size scheduler when performing guidance.</span>
                        </div>
                        <div class="tooltip">Final step size &#9432
                        <span class="tooltiptext">The final step size of the linear step size scheduler when performing guidance.</span>
                        </div>
                        <div class="tooltip">Number of self-attention clusters per subject &#9432
                        <span class="tooltiptext">Determines the number of clusters when clustering the self-attention maps (#clusters = #subject x #clusters_per_subject). Changing this value might improve semantics (adherence to the prompt), especially when the subjects exceed their bounding boxes.</span>
                        </div>
                        <div class="tooltip">Cross-attention loss scale factor &#9432
                        <span class="tooltiptext">The scale factor of the cross-attention loss term. Increasing it will improve semantic control (adherence to the prompt), but may reduce image quality.</span>
                        </div>
                        <div class="tooltip">Self-attention loss scale factor &#9432
                        <span class="tooltiptext">The scale factor of the self-attention loss term. Increasing it will improve layout control (adherence to the bounding boxes), but may reduce image quality.</span>
                        </div>
                        <div class="tooltip">Classifier-free guidance scale &#9432
                        <span class="tooltiptext">The scale factor of classifier-free guidance.</span>
                        </div>
                        <div class="tooltip" >Number of Gradient Descent iterations per timestep &#9432
                        <span class="tooltiptext">The number of Gradient Descent iterations for each timestep when performing guidance.</span>
                        </div>
                        <div class="tooltip" >Loss Threshold &#9432
                        <span class="tooltiptext">If the loss is below the threshold, Gradient Descent stops for that timestep. </span>
                        </div>
                        <div class="tooltip" >Number of guidance steps &#9432
                        <span class="tooltiptext">The number of timesteps in which to perform guidance.</span>
                        </div>
                    """
                    gr.HTML(description)
                    batch_size = gr.Slider(minimum=1, maximum=5, step=1, value=1, label="Number of samples")
                    init_step_size = gr.Slider(minimum=0, maximum=50, step=0.5, value=25, label="Initial step size")
                    final_step_size = gr.Slider(minimum=0, maximum=20, step=0.5, value=10, label="Final step size")
                    num_clusters_per_subject = gr.Slider(minimum=0, maximum=5, step=0.5, value=3, label="Number of clusters per subject")
                    cross_loss_scale = gr.Slider(minimum=0, maximum=2, step=0.1, value=1, label="Cross-attention loss scale factor")
                    self_loss_scale = gr.Slider(minimum=0, maximum=2, step=0.1, value=1, label="Self-attention loss scale factor")
                    classifier_free_guidance_scale = gr.Slider(minimum=0, maximum=50, step=0.5, value=7.5, label="Classifier-free guidance Scale")
                    num_iterations = gr.Slider(minimum=0, maximum=10, step=1, value=5, label="Number of Gradient Descent iterations")
                    loss_threshold = gr.Slider(minimum=0, maximum=1, step=0.1, value=0.2, label="Loss threshold")
                    num_guidance_steps = gr.Slider(minimum=10, maximum=20, step=1, value=15, label="Number of timesteps to perform guidance")
                    seed = gr.Slider(minimum=0, maximum=1000, step=1, value=445, label="Random Seed")

            boxes = gr.State([])

            sketch_pad.edit(
                draw,
                inputs=[boxes, sketch_pad, layout_image],
                outputs=[boxes, sketch_pad, layout_image],
                queue=False,
            )

            clear_button.click(
                clear,
                inputs=[batch_size],
                outputs=[boxes, sketch_pad, layout_image, out_images],
                queue=False,
            )

            generate_button.click(
                fn=partial(generate, device, model),
                inputs=[
                    prompt, subject_token_indices, filter_token_indices, num_tokens,
                    init_step_size, final_step_size, num_clusters_per_subject, cross_loss_scale, self_loss_scale,
                    classifier_free_guidance_scale, batch_size, num_iterations, loss_threshold, num_guidance_steps,
                    seed,
                    boxes,
                ],
                outputs=[out_images],
                queue=True,
            )

        #with gr.Column():
        #    gr.Examples(
        #        examples=[
        #            [
        #                [[0.35, 0.4, 0.65, 0.9], [0, 0.6, 0.3, 0.9], [0.7, 0.55, 1, 0.85]],
        #                "3D Pixar animation of a cute unicorn and a pink hedgehog and a nerdy owl traveling in a magical forest",
        #                "7,8,17;11,12,17;15,16,17",
        #                "5,6,9,10,13,14,18,19",
        #                286,
        #            ],
        #        ],
        #        inputs=[boxes, prompt, subject_token_indices, filter_token_indices, seed],
        #        outputs=None,
        #        fn=None,
        #        cache_examples=False,
        #    )
        description = """<p> The source code of this demo is based on the <a href="https://huggingface.co/spaces/gligen/demo/tree/main">GLIGEN demo</a>.</p>"""
        gr.HTML(description)

    demo.queue(max_size=50)
    demo.launch(share=False, show_api=False, show_error=True)

if __name__ == '__main__':
    main()