app.py

import gradio as gr

import torch
from src.euler_scheduler import MyEulerAncestralDiscreteScheduler
from diffusers.pipelines.auto_pipeline import AutoPipelineForImage2Image
from src.sdxl_inversion_pipeline import SDXLDDIMPipeline
from src.config import RunConfig
from src.editor import ImageEditorDemo

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

scheduler_class = MyEulerAncestralDiscreteScheduler

pipe_inversion = SDXLDDIMPipeline.from_pretrained("stabilityai/sdxl-turbo", use_safetensors=True).to(device)
pipe_inference = AutoPipelineForImage2Image.from_pretrained("stabilityai/sdxl-turbo", use_safetensors=True).to(device)
pipe_inference.scheduler = scheduler_class.from_config(pipe_inference.scheduler.config)
pipe_inversion.scheduler = scheduler_class.from_config(pipe_inversion.scheduler.config)
pipe_inversion.scheduler_inference = scheduler_class.from_config(pipe_inference.scheduler.config)


# if torch.cuda.is_available():
#     torch.cuda.max_memory_allocated(device=device)
#     pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16", use_safetensors=True)
#     pipe.enable_xformers_memory_efficient_attention()
#     pipe = pipe.to(device)
# else: 
#     pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", use_safetensors=True)
#     pipe = pipe.to(device)


def set_pipe(input_image, description_prompt, edit_guidance_scale, num_inference_steps=4,
          num_inversion_steps=4, inversion_max_step=0.6):
    config = RunConfig(num_inference_steps=num_inference_steps,
                       num_inversion_steps=num_inversion_steps,
                       edit_guidance_scale=edit_guidance_scale,
                       inversion_max_step=inversion_max_step)

    return ImageEditorDemo(pipe_inversion, pipe_inference, input_image, description_prompt, config, device)

    # return editor
    # image = editor.edit(target_prompt)
    # return image

def edit(editor, target_prompt):
    if editor is None:
        return "Please set image & description before editing."
    return editor.edit(target_prompt)

examples = [
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    "An astronaut riding a green horse",
    "A delicious ceviche cheesecake slice",
]

# css = """
# #col-container-1 {
#     margin: 0 auto;
#     max-width: 520px;
# }
# #col-container-2 {
#     margin: 0 auto;
#     max-width: 520px;
# }
# """

if torch.cuda.is_available():
    power_device = "GPU"
else:
    power_device = "CPU"

# with gr.Blocks(css=css) as demo:
with gr.Blocks() as demo:
    gr.Markdown(f"""
    This is a demo for our [paper](https://arxiv.org/abs/2312.12540) **RNRI: Regularized Newton Raphson Inversion for Text-to-Image Diffusion Models**.
    Image editing using our RNRI for inversion demonstrates significant speed-up and improved quality compared to previous state-of-the-art methods.
    Take a look at our [project page](https://barakmam.github.io/rnri.github.io/).
    """)
    with gr.Row():
        with gr.Column(elem_id="col-container-1"):
            with gr.Row():
                input_image = gr.Image(label="Input image", sources=['upload', 'webcam'], type="pil")

            with gr.Row():
                description_prompt = gr.Text(
                    label="Image description",
                    show_label=False,
                    max_lines=1,
                    placeholder="Enter your image description",
                    container=False,
                )

            with gr.Row():
                set_button = gr.Button("Set input image & description", scale=1)

            editor_state = gr.State()
            with gr.Row():
                target_prompt = gr.Text(
                    label="Edit prompt",
                    show_label=False,
                    max_lines=1,
                    placeholder="Enter your edit prompt",
                    container=False,
                )

            with gr.Accordion("Advanced Settings", open=False):
                with gr.Row():
                    edit_guidance_scale = gr.Slider(
                        label="Guidance scale",
                        minimum=0.0,
                        maximum=10.0,
                        step=0.1,
                        value=1.2,
                    )

                    num_inference_steps = gr.Slider(
                        label="Number of RNRI iterations",
                        minimum=1,
                        maximum=12,
                        step=1,
                        value=4,
                    )

                    inversion_max_step = gr.Slider(
                        label="Inversion strength",
                        minimum=0.0,
                        maximum=1.0,
                        step=0.01,
                        value=0.6,
                    )

            with gr.Row():
                run_button = gr.Button("Edit", scale=1)

        with gr.Column(elem_id="col-container-2"):
            result = gr.Image(label="Result")

            # gr.Examples(
            #     examples = examples,
            #     inputs = [prompt]
            # )

    set_button.click(
        fn=set_pipe,
        inputs=[input_image, description_prompt, edit_guidance_scale, num_inference_steps,
                num_inference_steps, inversion_max_step],
        outputs=editor_state
    )

    run_button.click(
        fn=edit,
        inputs=[editor_state, target_prompt],
        outputs=[result]
    )


demo.queue().launch()

# im = infer(input_image, description_prompt, target_prompt, edit_guidance_scale, num_inference_steps=4, num_inversion_steps=4,
#           inversion_max_step=0.6)