import math
import random
import gradio as gr
import torch
from PIL import Image, ImageOps
from diffusers import StableDiffusionInstructPix2PixPipeline
import spaces
help_text = """
Considerations while editing:
1. The Base-Model, trained on the PIPE dataset, is great for some tasks, while the Finetuned-MB-Model, fine-tuned on the MagicBrush dataset, can be better for others. Please try both until you are satisfied.
2. Image CFG controls how much to deviate from the original image. Higher values keep the image more consistent with the original.
3. Text CFG does the opposite. Higher values lead to more changes in the image.
4. Using different seed values will produce varied outputs.
5. Increasing the number of steps can enhance the results.
6. The Stable Diffusion autoencoder struggles with small faces in images.
"""
article = """
Paint by Inpaint: Learning to Add Image Objects by Removing Them First
"""
description = """
Gradio demo for Paint by Inpaint: Learning to Add Image Objects by Removing Them First, visit our project page.
The demo involves two models: one trained for image object addition using the PIPE dataset, and another model further fine-tuned on the MagicBrush dataset.
"""
# Base models
object_addition_base_model_id = "paint-by-inpaint/add-base"
# general_editing_base_model_id = "paint-by-inpaint/general-base"
# MagicBrush finetuned models
object_addition_finetuned_model_id = "paint-by-inpaint/add-finetuned-mb"
# general_editing_finetuned_model_id = "paint-by-inpaint/general-finetuned-mb"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if "cuda" in device else torch.float32
def load_model(model_id):
return StableDiffusionInstructPix2PixPipeline.from_pretrained(model_id, torch_dtype=dtype).to(device)
pipe_object_addition_base = load_model(object_addition_base_model_id)
pipe_object_addition_finetuned = load_model(object_addition_finetuned_model_id)
# pipe_general_editing_base = load_model(general_editing_base_model_id)
# pipe_general_editing_finetuned = load_model(general_editing_finetuned_model_id)
@spaces.GPU(duration=15)
def generate(
input_image: Image.Image,
instruction: str,
model_choice: int,
steps: int,
randomize_seed: bool,
seed: int,
text_cfg_scale: float,
image_cfg_scale: float,
task_type: str,
):
seed = random.randint(0, 100000) if randomize_seed else seed
if task_type == "object_addition":
pipe = pipe_object_addition_base if model_choice == 0 else pipe_object_addition_finetuned
else:
pipe = pipe_general_editing_base if model_choice == 0 else pipe_general_editing_finetuned
width, height = input_image.size
factor = 512 / max(width, height)
factor = math.ceil(min(width, height) * factor / 64) * 64 / min(width, height)
width = int((width * factor) // 64) * 64
height = int((height * factor) // 64) * 64
input_image = ImageOps.fit(input_image, (width, height), method=Image.Resampling.LANCZOS)
if instruction == "":
return [input_image, seed]
generator = torch.manual_seed(seed)
edited_image = pipe(
instruction, image=input_image,
guidance_scale=text_cfg_scale, image_guidance_scale=image_cfg_scale,
num_inference_steps=steps, generator=generator,
).images[0]
return [seed, text_cfg_scale, image_cfg_scale, edited_image]
def reset():
return [0, "Randomize Seed", 2024, "Fix CFG", 7.5, 1.5, None]
with gr.Blocks(css=".compact-box .gr-row { margin-bottom: 5px; } .compact-box .gr-number input, .compact-box .gr-radio label { padding: 5px 10px; }") as demo:
gr.HTML("""
Paint by Inpaint
{description}
""".format(description=description))
# with gr.Tabs():
# with gr.Tab("Object Addition"):
if 1:
with gr.Row():
with gr.Column():
input_image = gr.Image(label="Input Image", type="pil", interactive=True)
instruction = gr.Textbox(lines=1, label="Addition Instruction", interactive=True, max_lines=1, placeholder="Enter addition instruction here")
model_choice = gr.Radio(
["Base-Model", "Finetuned-MB-Model"],
value="Base-Model",
type="index",
label="Choose Model",
interactive=True,
)
with gr.Group(elem_id="compact-box"):
with gr.Row():
with gr.Column():
steps = gr.Number(value=50, precision=0, label="Steps", interactive=True)
with gr.Column():
with gr.Row():
seed = gr.Number(value=2024, precision=0, label="Seed", interactive=True)
randomize_seed = gr.Radio(
["Fix Seed", "Randomize Seed"],
value="Randomize Seed",
type="index",
show_label=False,
interactive=True,
)
with gr.Row():
text_cfg_scale = gr.Number(value=7.5, label="Text CFG", interactive=True)
image_cfg_scale = gr.Number(value=1.5, label="Image CFG", interactive=True)
with gr.Row():
generate_button = gr.Button("Generate")
reset_button = gr.Button("Reset")
with gr.Column():
edited_image = gr.Image(label="Edited Image", type="pil", interactive=False)
generate_button.click(
fn=lambda *args: generate(*args, task_type="object_addition"),
inputs=[
input_image,
instruction,
model_choice,
steps,
randomize_seed,
seed,
text_cfg_scale,
image_cfg_scale,
],
outputs=[seed, text_cfg_scale, image_cfg_scale, edited_image],
)
reset_button.click(
fn=reset,
inputs=[],
outputs=[steps, randomize_seed, seed, text_cfg_scale, image_cfg_scale, edited_image],
)
# with gr.Tab("General Editing"):
# with gr.Row():
# with gr.Column():
# input_image_editing = gr.Image(label="Input Image", type="pil", interactive=True)
# instruction_editing = gr.Textbox(lines=1, label="Editing Instruction", interactive=True, max_lines=1, placeholder="Enter editing instruction here")
# model_choice_editing = gr.Radio(
# ["Base-Model", "Finetuned-MB-Model"],
# value="Base-Model",
# type="index",
# label="Choose Model",
# interactive=True,
# )
# with gr.Group(elem_id="compact-box"):
# with gr.Row():
# steps_editing = gr.Number(value=50, precision=0, label="Steps", interactive=True)
# with gr.Column():
# with gr.Row():
# seed_editing = gr.Number(value=2024, precision=0, label="Seed", interactive=True)
# randomize_seed_editing = gr.Radio(
# ["Fix Seed", "Randomize Seed"],
# value="Randomize Seed",
# type="index",
# show_label=False,
# interactive=True,
# )
# with gr.Row():
# text_cfg_scale_editing = gr.Number(value=7.5, label="Text CFG", interactive=True)
# image_cfg_scale_editing = gr.Number(value=1.5, label="Image CFG", interactive=True)
# with gr.Row():
# generate_button_editing = gr.Button("Generate")
# reset_button_editing = gr.Button("Reset")
# with gr.Column():
# edited_image_editing = gr.Image(label="Edited Image", type="pil", interactive=False)
# generate_button_editing.click(
# fn=lambda *args: generate(*args, task_type="general_editing"),
# inputs=[
# input_image_editing,
# instruction_editing,
# model_choice_editing,
# steps_editing,
# randomize_seed_editing,
# seed_editing,
# text_cfg_scale_editing,
# image_cfg_scale_editing,
# ],
# outputs=[seed_editing, text_cfg_scale_editing, image_cfg_scale_editing, edited_image_editing],
# )
# reset_button_editing.click(
# fn=reset,
# inputs=[],
# outputs=[steps_editing, randomize_seed_editing, seed_editing, text_cfg_scale_editing, image_cfg_scale_editing, edited_image_editing],
# )
gr.Markdown(help_text)
examples = [
["examples/messi.jpeg", "Add a royal silver crown"],
["examples/coffee.jpg", "Add steamed milk"],
]
gr.Examples(
examples=examples,
inputs=[input_image, instruction],
outputs=[edited_image],
)
gr.HTML(article)
demo.queue()
demo.launch(share=False, max_threads=1)