Spaces:
Runtime error
Runtime error
File size: 8,494 Bytes
04bf3ab 3796cef 04bf3ab 8cd1abb 7d008e4 04bf3ab 3e7b7cc 04bf3ab a833338 b4542eb d16d053 04bf3ab d16d053 04bf3ab a833338 04bf3ab d16d053 04bf3ab d16d053 7d008e4 04bf3ab d16d053 04bf3ab d16d053 557cf2f 7d008e4 557cf2f 7d008e4 557cf2f d16d053 7d008e4 5d57acf 7d008e4 5d57acf 7d008e4 d16d053 8cd1abb d16d053 7d008e4 d16d053 7d008e4 8cd1abb 7d008e4 d16d053 7d008e4 557cf2f 693d97d 557cf2f 693d97d d16d053 8cd1abb 7d008e4 d16d053 7d008e4 8cd1abb 557cf2f d16d053 7d008e4 8cd1abb 5d57acf 557cf2f 8cd1abb 7d008e4 d16d053 7d008e4 8cd1abb 7d008e4 8cd1abb 7d008e4 693d97d 5d57acf 8cd1abb 7d008e4 d16d053 8cd1abb 5d57acf 7d008e4 d16d053 8cd1abb 557cf2f d16d053 3cc08ec 5d57acf 8cd1abb d16d053 8cd1abb 693d97d 557cf2f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 |
import io
import requests
import numpy as np
import torch
import os
from PIL import Image
from typing import List, Optional
from functools import reduce
import gradio as gr
from transformers import DetrFeatureExtractor, DetrForSegmentation, DetrConfig
from transformers.models.detr.feature_extraction_detr import rgb_to_id
from diffusers import StableDiffusionInpaintPipeline
auth_token = os.environ.get("READ_TOKEN")
torch.inference_mode()
torch.no_grad()
# Load segmentation models
def load_segmentation_models(model_name: str = 'facebook/detr-resnet-50-panoptic'):
feature_extractor = DetrFeatureExtractor.from_pretrained(model_name)
model = DetrForSegmentation.from_pretrained(model_name)
cfg = DetrConfig.from_pretrained(model_name)
return feature_extractor, model, cfg
# Load diffusion pipeline
def load_diffusion_pipeline(model_name: str = 'runwayml/stable-diffusion-inpainting'):
return StableDiffusionInpaintPipeline.from_pretrained(
model_name,
revision='fp16',
torch_dtype=torch.float16,
use_auth_token=auth_token
)
# Device helper
def get_device(try_cuda=True):
return torch.device('cuda' if try_cuda and torch.cuda.is_available() else 'cpu')
def min_pool(x: torch.Tensor, kernel_size: int):
pad_size = (kernel_size - 1) // 2
return -torch.nn.functional.max_pool2d(-x, kernel_size, (1, 1), padding=pad_size)
def max_pool(x: torch.Tensor, kernel_size: int):
pad_size = (kernel_size - 1) // 2
return torch.nn.functional.max_pool2d(x, kernel_size, (1, 1), padding=pad_size)
# Apply min-max pooling to clean up mask
def clean_mask(mask, max_kernel: int = 23, min_kernel: int = 5):
mask = torch.Tensor(mask[None, None]).float()
mask = min_pool(mask, min_kernel)
mask = max_pool(mask, max_kernel)
mask = mask.bool().squeeze().numpy()
return mask
feature_extractor, segmentation_model, segmentation_cfg = load_segmentation_models()
pipe = load_diffusion_pipeline()
device = get_device()
pipe = pipe.to(device)
# Callback function that runs segmentation and updates CheckboxGroup
def fn_segmentation(image, max_kernel, min_kernel):
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = segmentation_model(**inputs)
processed_sizes = torch.as_tensor(inputs["pixel_values"].shape[-2:]).unsqueeze(0)
result = feature_extractor.post_process_panoptic(outputs, processed_sizes)[0]
panoptic_seg = Image.open(io.BytesIO(result["png_string"])).resize((image.width, image.height))
panoptic_seg = np.array(panoptic_seg, dtype=np.uint8)
panoptic_seg_id = rgb_to_id(panoptic_seg)
raw_masks = []
for s in result['segments_info']:
m = panoptic_seg_id == s['id']
raw_masks.append(m.astype(np.uint8) * 255)
checkbox_choices = [f"{s['id']}:{segmentation_cfg.id2label[s['category_id']]}" for s in result['segments_info']]
checkbox_group = gr.CheckboxGroup.update(
choices=checkbox_choices
)
return raw_masks, checkbox_group, gr.Image.update(value=np.zeros((image.height, image.width))), gr.Image.update(value=image)
# Callback function that updates the displayed mask based on selected checkboxes
def fn_update_mask(
image: Image,
masks: List[np.array],
masks_enabled: List[int],
max_kernel: int,
min_kernel: int,
invert_mask: bool
):
masks_enabled = [int(m.split(':')[0]) for m in masks_enabled]
combined_mask = reduce(lambda x, y: x | y, [masks[i] for i in masks_enabled], np.zeros_like(masks[0], dtype=bool))
if invert_mask:
combined_mask = ~combined_mask
combined_mask = clean_mask(combined_mask, max_kernel, min_kernel)
masked_image = np.array(image).copy()
masked_image[combined_mask] = 0.0
return combined_mask.astype(np.uint8) * 255, Image.fromarray(masked_image)
# Callback function that runs diffusion given the current image, mask and prompt.
def fn_diffusion(
prompt: str,
masked_image: Image,
mask: Image,
num_diffusion_steps: int,
guidance_scale: float,
negative_prompt: Optional[str] = None,
):
if len(negative_prompt) == 0:
negative_prompt = None
# Resize image to a more stable diffusion friendly format.
# TODO: remove magic number
STABLE_DIFFUSION_SMALL_EDGE = 512
w, h = masked_image.size
is_width_larger = w > h
resize_ratio = STABLE_DIFFUSION_SMALL_EDGE / (h if is_width_larger else w)
new_width = int(w * resize_ratio) if is_width_larger else STABLE_DIFFUSION_SMALL_EDGE
new_height = STABLE_DIFFUSION_SMALL_EDGE if is_width_larger else int(h * resize_ratio)
new_width += 8 - (new_width % 8) if is_width_larger else 0
new_height += 0 if is_width_larger else 8 - (new_height % 8)
mask = Image.fromarray(mask).convert("RGB").resize((new_width, new_height))
masked_image = masked_image.convert("RGB").resize((new_width, new_height))
# Run diffusion
inpainted_image = pipe(
height=new_height,
width=new_width,
prompt=prompt,
image=masked_image,
mask_image=mask,
num_inference_steps=num_diffusion_steps,
guidance_scale=guidance_scale,
negative_prompt=negative_prompt
).images[0]
# Resize back to the original size
inpainted_image = inpainted_image.resize((w, h))
return inpainted_image
demo = gr.Blocks(css=open('app.css').read())
with demo:
gr.HTML(open('app_header.html').read())
# gr.Markdown("# Stable Diffusion Inpainting Segmentation")
# Input image control
input_image = gr.Image(value="http://images.cocodataset.org/val2017/000000039769.jpg", type='pil', label="Input Image")
# Combined mask controls
bt_masks = gr.Button("Compute Masks")
with gr.Row():
mask_image = gr.Image(type='numpy', label="Diffusion Mask")
masked_image = gr.Image(type='pil', label="Masked Image")
mask_storage = gr.State()
# Mask editing controls
with gr.Row():
max_slider = gr.Slider(minimum=1, maximum=99, value=23, step=2, label="Mask Overflow")
min_slider = gr.Slider(minimum=1, maximum=99, value=5, step=2, label="Mask Denoising")
invert_mask = gr.Checkbox(label="Invert Mask")
mask_checkboxes = gr.CheckboxGroup(interactive=True, label="Mask Selection")
# Diffusion controls and output
with gr.Row():
with gr.Column():
prompt = gr.Textbox("Two ginger cats lying together on a pink sofa. There are two TV remotes. High definition.", label="Prompt")
negative_prompt = gr.Textbox(label="Negative Prompt")
with gr.Column():
steps_slider = gr.Slider(minimum=1, maximum=100, value=50, label="Inference Steps")
guidance_slider = gr.Slider(minimum=0.0, maximum=50.0, value=7.5, step=0.1, label="Guidance Scale")
bt_diffusion = gr.Button("Run Diffusion")
inpainted_image = gr.Image(type='pil', label="Inpainted Image")
# TODO: saw a better way of handling many inputs online..
# forgot where though
update_mask_inputs = [input_image, mask_storage, mask_checkboxes, max_slider, min_slider, invert_mask]
update_mask_outputs = [mask_image, masked_image]
# Clear checkbox group on input image change
input_image.change(lambda: gr.CheckboxGroup.update(choices=[], value=[]), outputs=mask_checkboxes)
input_image.change(lambda: gr.Checkbox.update(value=False), outputs=invert_mask)
# Segmentation button callback
bt_masks.click(fn_segmentation, inputs=[input_image, max_slider, min_slider], outputs=[mask_storage, mask_checkboxes, mask_image, masked_image])
# Update mask callbacks
max_slider.change(fn_update_mask, inputs=update_mask_inputs, outputs=update_mask_outputs, show_progress=False)
min_slider.change(fn_update_mask, inputs=update_mask_inputs, outputs=update_mask_outputs, show_progress=False)
mask_checkboxes.change(fn_update_mask, inputs=update_mask_inputs, outputs=update_mask_outputs, show_progress=False)
invert_mask.change(fn_update_mask, inputs=update_mask_inputs, outputs=update_mask_outputs, show_progress=False)
# Diffusion button callback
bt_diffusion.click(fn_diffusion, inputs=[
prompt,
masked_image,
mask_image,
steps_slider,
guidance_slider,
negative_prompt
], outputs=inpainted_image)
gr.HTML(open('app_license.html').read())
demo.launch()
|