app.py

import gradio as gr
import spaces
import torch

from diffusers import AutoencoderKL, ControlNetUnionModel, DiffusionPipeline, TCDScheduler


def callback_cfg_cutoff(pipeline, step_index, timestep, callback_kwargs):
    if step_index == int(pipeline.num_timesteps * 0.2):
        prompt_embeds = callback_kwargs["prompt_embeds"]
        prompt_embeds = prompt_embeds[-1:]

        add_text_embeds = callback_kwargs["add_text_embeds"]
        add_text_embeds = add_text_embeds[-1:]

        add_time_ids = callback_kwargs["add_time_ids"]
        add_time_ids = add_time_ids[-1:]

        control_image = callback_kwargs["control_image"]
        control_image[0] = control_image[0][-1:]

        control_type = callback_kwargs["control_type"]
        control_type = control_type[-1:]

        pipeline._guidance_scale = 0.0
        callback_kwargs["prompt_embeds"] = prompt_embeds
        callback_kwargs["add_text_embeds"] = add_text_embeds
        callback_kwargs["add_time_ids"] = add_time_ids
        callback_kwargs["control_image"] = control_image
        callback_kwargs["control_type"] = control_type

    return callback_kwargs


MODELS = {
    "RealVisXL V5.0 Lightning": "SG161222/RealVisXL_V5.0_Lightning",
}

controlnet_model = ControlNetUnionModel.from_pretrained(
    "OzzyGT/controlnet-union-promax-sdxl-1.0", variant="fp16", torch_dtype=torch.float16
)
controlnet_model.to(device="cuda", dtype=torch.float16)
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to("cuda")

pipe = DiffusionPipeline.from_pretrained(
    "SG161222/RealVisXL_V5.0_Lightning",
    torch_dtype=torch.float16,
    vae=vae,
    controlnet=controlnet_model,
    custom_pipeline="OzzyGT/custom_sdxl_cnet_union",
).to("cuda")

pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)


@spaces.GPU(duration=24)
def fill_image(prompt, negative_prompt, image, model_selection, paste_back):
    (
        prompt_embeds,
        negative_prompt_embeds,
        pooled_prompt_embeds,
        negative_pooled_prompt_embeds,
    ) = pipe.encode_prompt(prompt, device="cuda", negative_prompt=negative_prompt)

    source = image["background"]
    mask = image["layers"][0]

    alpha_channel = mask.split()[3]
    binary_mask = alpha_channel.point(lambda p: p > 0 and 255)
    cnet_image = source.copy()
    cnet_image.paste(0, (0, 0), binary_mask)

    image = pipe(
        prompt_embeds=prompt_embeds,
        negative_prompt_embeds=negative_prompt_embeds,
        pooled_prompt_embeds=pooled_prompt_embeds,
        negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
        control_image=[cnet_image],
        controlnet_conditioning_scale=[1.0],
        control_mode=[7],
        num_inference_steps=8,
        guidance_scale=1.5,
        callback_on_step_end=callback_cfg_cutoff,
        callback_on_step_end_tensor_inputs=[
            "prompt_embeds",
            "add_text_embeds",
            "add_time_ids",
            "control_image",
            "control_type",
        ],
    ).images[0]

    if paste_back:
        image = image.convert("RGBA")
        cnet_image.paste(image, (0, 0), binary_mask)
    else:
        cnet_image = image

    yield source, cnet_image


def clear_result():
    return gr.update(value=None)


title = """<h2 align="center">Diffusers Fast Inpaint</h2>
<div align="center">Draw the mask over the subject you want to erase or change and write what you want to inpaint it with.</div>
"""

with gr.Blocks() as demo:
    gr.HTML(title)
    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(
                label="Prompt",
                lines=1,
            )
        with gr.Column():
            with gr.Row():
                negative_prompt = gr.Textbox(
                    label="Negative Prompt",
                    lines=1,
                )

    with gr.Row():
        with gr.Column():
            run_button = gr.Button("Generate")

        with gr.Column():
            paste_back = gr.Checkbox(True, label="Paste back original")

    with gr.Row():
        input_image = gr.ImageMask(
            type="pil",
            label="Input Image",
            crop_size=(1024, 1024),
            canvas_size=(1024, 1024),
            layers=False,
            height=512,
        )

        result = gr.ImageSlider(
            interactive=False,
            label="Generated Image",
        )

    use_as_input_button = gr.Button("Use as Input Image", visible=False)

    model_selection = gr.Dropdown(choices=list(MODELS.keys()), value="RealVisXL V5.0 Lightning", label="Model")

    def use_output_as_input(output_image):
        return gr.update(value=output_image[1])

    use_as_input_button.click(fn=use_output_as_input, inputs=[result], outputs=[input_image])

    run_button.click(
        fn=clear_result,
        inputs=None,
        outputs=result,
    ).then(
        fn=lambda: gr.update(visible=False),
        inputs=None,
        outputs=use_as_input_button,
    ).then(
        fn=fill_image,
        inputs=[prompt, negative_prompt, input_image, model_selection, paste_back],
        outputs=result,
    ).then(
        fn=lambda: gr.update(visible=True),
        inputs=None,
        outputs=use_as_input_button,
    )

    prompt.submit(
        fn=clear_result,
        inputs=None,
        outputs=result,
    ).then(
        fn=lambda: gr.update(visible=False),
        inputs=None,
        outputs=use_as_input_button,
    ).then(
        fn=fill_image,
        inputs=[prompt, negative_prompt, input_image, model_selection, paste_back],
        outputs=result,
    ).then(
        fn=lambda: gr.update(visible=True),
        inputs=None,
        outputs=use_as_input_button,
    )


demo.queue(max_size=12).launch(share=False)