|
import os |
|
import gradio as gr |
|
import torch |
|
import yaml |
|
import numpy as np |
|
|
|
from munch import munchify |
|
import torchvision.transforms as transforms |
|
from torchvision.transforms import functional as F |
|
from diffusers import ( |
|
AutoPipelineForInpainting, |
|
) |
|
from generate_dataset import outpainting_generator_rectangle |
|
|
|
|
|
transform = transforms.Compose([ |
|
transforms.ToPILImage(), |
|
transforms.Resize((512, 512), interpolation=F.InterpolationMode.LANCZOS), |
|
]) |
|
|
|
def pref_inpainting(image, |
|
box_width_ratio, |
|
mask_random_start, |
|
steps, |
|
): |
|
with open("/data0/kendong/Diffusions/zero123-live/configs/imagereward_train_configs.yaml") as file: |
|
config_dict= yaml.safe_load(file) |
|
config = munchify(config_dict) |
|
|
|
pipe = AutoPipelineForInpainting.from_pretrained( |
|
'/data1/kendong/joint-rl-diffusion/alignment_log/exp_reward_group_regression_all_1w_1.6boundary/iteration_2560', num_inference_steps=steps) |
|
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") |
|
pipe = pipe.to(device) |
|
|
|
|
|
color, mask = outpainting_generator_rectangle(image, box_width_ratio/100, mask_random_start) |
|
mask = mask.convert('L') |
|
|
|
color, mask = np.array(color).transpose(2, 0, 1), np.array(mask) |
|
mask = mask[None, ...] |
|
mask_ = np.zeros_like(mask) |
|
mask_[mask < 125] = 0 |
|
mask_[mask >= 125] = 1 |
|
|
|
color = torch.from_numpy(color).to(device) |
|
mask = torch.from_numpy(mask).to(device) |
|
|
|
|
|
color, mask = transform(color), transform(mask) |
|
res = pipe(prompt='', image=color, mask_image=mask, eta=config.eta).images[0] |
|
|
|
|
|
|
|
|
|
return color, res |
|
|
|
|
|
inputs = [ |
|
gr.Image(type="pil", image_mode="RGBA", label='Input Image'), |
|
gr.Slider(30, 45, value=35, step=1, label="box_width_ratio"), |
|
gr.Slider(0, 256, value=125, step=1, label="mask_random_start"), |
|
gr.Slider(30, 100, value=50, step=5, label="steps"), |
|
] |
|
|
|
outputs = [ |
|
gr.Image(type="pil", image_mode="RGBA", label='Mask RGB Image', container=True, width="65%"), |
|
gr.Image(type="pil", image_mode="RGBA", label='Results', container=True, width="65%"), |
|
] |
|
|
|
examples = [ |
|
["/data0/kendong/Diffusions/zero123-live/test_demo/assets/ILSVRC2012_test_00000003.JPEG", 35, 125, 50], |
|
["/data0/kendong/Diffusions/zero123-live/test_demo/assets/ILSVRC2012_test_00000181.JPEG", 35, 125, 50], |
|
["/data0/kendong/Diffusions/zero123-live/test_demo/assets/ILSVRC2012_test_00002334.JPEG", 35, 125, 50], |
|
["/data0/kendong/Diffusions/zero123-live/test_demo/assets/ILSVRC2012_test_00002613.JPEG", 35, 125, 50], |
|
] |
|
|
|
|
|
|
|
iface = gr.Interface( |
|
fn=pref_inpainting, |
|
inputs=inputs, |
|
outputs=outputs, |
|
title="Inpainting with Human Preference (Utilizing Free CPU Resources)", |
|
description="Upload an image and start your inpainting (currently only supporting outpainting masks; other mask types coming soon).", |
|
theme="default", |
|
examples= examples, |
|
allow_flagging="never" |
|
) |
|
|
|
|
|
iface.launch(share=True) |
|
|
