sd-webui-segment-anything/scripts/sam.py

443 lines
23 KiB
Python

import gc
import os
import copy
import glob
import numpy as np
from PIL import Image
import torch
import gradio as gr
from collections import OrderedDict
from scipy.ndimage import binary_dilation
from modules import scripts, shared
from modules.ui import gr_show
from modules.safe import unsafe_torch_load, load
from modules.processing import StableDiffusionProcessingImg2Img
from modules.devices import device, torch_gc, cpu
from modules.paths import models_path
from segment_anything import SamPredictor, sam_model_registry
from scripts.dino import dino_model_list, dino_predict_internal, show_boxes, clear_dino_cache
sam_model_cache = OrderedDict()
scripts_sam_model_dir = os.path.join(scripts.basedir(), "models/sam")
sd_sam_model_dir = os.path.join(models_path, "sam")
sam_model_dir = sd_sam_model_dir if os.path.exists(sd_sam_model_dir) else scripts_sam_model_dir
sam_model_list = [f for f in os.listdir(sam_model_dir) if os.path.isfile(
os.path.join(sam_model_dir, f)) and f.split('.')[-1] != 'txt']
refresh_symbol = '\U0001f504' # 🔄
class ToolButton(gr.Button, gr.components.FormComponent):
"""Small button with single emoji as text, fits inside gradio forms"""
def __init__(self, **kwargs):
super().__init__(variant="tool", **kwargs)
def get_block_name(self):
return "button"
def show_masks(image_np, masks: np.ndarray, alpha=0.5):
image = copy.deepcopy(image_np)
np.random.seed(0)
for mask in masks:
color = np.concatenate([np.random.random(3), np.array([0.6])], axis=0)
image[mask] = image[mask] * (1 - alpha) + 255 * color.reshape(1, 1, -1) * alpha
return image.astype(np.uint8)
def load_sam_model(sam_checkpoint):
model_type = '_'.join(sam_checkpoint.split('_')[1:-1])
sam_checkpoint = os.path.join(sam_model_dir, sam_checkpoint)
torch.load = unsafe_torch_load
sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)
sam.to(device=device)
sam.eval()
torch.load = load
return sam
def clear_sam_cache():
sam_model_cache.clear()
gc.collect()
torch_gc()
def update_mask(mask_gallery, chosen_mask, dilation_amt, input_image):
print("Dilation Amount: ", dilation_amt)
mask_image = Image.open(mask_gallery[chosen_mask + 3]['name'])
binary_img = np.array(mask_image.convert('1'))
if dilation_amt:
# Convert the image to a binary numpy array
mask_image, binary_img = dilate_mask(binary_img, dilation_amt)
blended_image = Image.fromarray(show_masks(np.array(input_image), binary_img.astype(np.bool_)[None, ...]))
matted_image = np.array(input_image)
matted_image[~binary_img] = np.array([0, 0, 0, 0])
return [blended_image, mask_image, Image.fromarray(matted_image)]
def clear_cache():
clear_sam_cache()
clear_dino_cache()
def refresh_sam_models(*inputs):
global sam_model_list
sam_model_list = [f for f in os.listdir(sam_model_dir) if os.path.isfile(
os.path.join(sam_model_dir, f)) and f.split('.')[-1] != 'txt']
dd = inputs[0]
if dd in sam_model_list:
selected = dd
elif len(sam_model_list) > 0:
selected = sam_model_list[0]
else:
selected = None
return gr.Dropdown.update(choices=sam_model_list, value=selected)
def dilate_mask(mask, dilation_amt):
# Create a dilation kernel
x, y = np.meshgrid(np.arange(dilation_amt), np.arange(dilation_amt))
center = dilation_amt // 2
dilation_kernel = ((x - center)**2 + (y - center)**2 <= center**2).astype(np.uint8)
# Dilate the image
dilated_binary_img = binary_dilation(mask, dilation_kernel)
# Convert the dilated binary numpy array back to a PIL image
dilated_mask = Image.fromarray(dilated_binary_img.astype(np.uint8) * 255)
return dilated_mask, dilated_binary_img
def init_sam_model(sam_model_name):
print("Initializing SAM")
if sam_model_name in sam_model_cache:
sam = sam_model_cache[sam_model_name]
if shared.cmd_opts.lowvram:
sam.to(device=device)
return sam
elif sam_model_name in sam_model_list:
clear_sam_cache()
sam_model_cache[sam_model_name] = load_sam_model(sam_model_name)
return sam_model_cache[sam_model_name]
else:
Exception(
f"{sam_model_name} not found, please download model to models/sam.")
def sam_predict(sam_model_name, input_image, positive_points, negative_points,
dino_checkbox, dino_model_name, text_prompt, box_threshold,
dino_preview_checkbox, dino_preview_boxes_selection, gui=True):
print("Start SAM Processing")
image_np = np.array(input_image)
image_np_rgb = image_np[..., :3]
dino_enabled = dino_checkbox and text_prompt is not None
boxes_filt = None
sam_predict_result = " done."
if dino_enabled:
boxes_filt, install_success = dino_predict_internal(input_image, dino_model_name, text_prompt, box_threshold)
if install_success and dino_preview_checkbox is not None and dino_preview_checkbox and dino_preview_boxes_selection is not None:
valid_indices = [int(i) for i in dino_preview_boxes_selection if int(i) < boxes_filt.shape[0]]
boxes_filt = boxes_filt[valid_indices]
if not install_success:
if len(positive_points) == 0 and len(negative_points) == 0:
return [], "GroundingDINO installment has failed. Check your terminal for more detail and submit an issue to https://github.com/continue-revolution/sd-webui-segment-anything/issues."
else:
sam_predict_result += " However, GroundingDINO installment has failed. Your process automatically fall back to point prompt only. Check your terminal for more detail and submit an issue to https://github.com/continue-revolution/sd-webui-segment-anything/issues."
sam = init_sam_model(sam_model_name)
print(f"Running SAM Inference {image_np_rgb.shape}")
predictor = SamPredictor(sam)
predictor.set_image(image_np_rgb)
if dino_enabled and boxes_filt.shape[0] > 1:
sam_predict_status = f"SAM inference with {boxes_filt.shape[0]} boxes, point prompts disgarded"
print(sam_predict_status)
transformed_boxes = predictor.transform.apply_boxes_torch(boxes_filt, image_np.shape[:2])
masks, _, _ = predictor.predict_torch(
point_coords=None,
point_labels=None,
boxes=transformed_boxes.to(device),
multimask_output=True,
)
masks = masks.permute(1, 0, 2, 3).cpu().numpy()
else:
sam_predict_status = f"SAM inference with {0 if boxes_filt is None else boxes_filt.shape[0]} box, {len(positive_points)} positive prompts, {len(negative_points)} negative prompts"
print(sam_predict_status)
point_coords = np.array(positive_points + negative_points)
point_labels = np.array([1] * len(positive_points) + [0] * len(negative_points))
box = copy.deepcopy(boxes_filt[0].numpy()) if boxes_filt is not None and boxes_filt.shape[0] > 0 else None
masks, _, _ = predictor.predict(
point_coords=point_coords if len(point_coords) > 0 else None,
point_labels=point_labels if len(point_coords) > 0 else None,
box=box,
multimask_output=True,
)
masks = masks[:, None, ...]
if shared.cmd_opts.lowvram:
sam.to(cpu)
gc.collect()
torch_gc()
print("Creating output image")
mask_images = []
masks_gallery = []
matted_images = []
boxes_filt = boxes_filt.numpy().astype(int) if boxes_filt is not None else None
for mask in masks:
masks_gallery.append(Image.fromarray(np.any(mask, axis=0)))
if gui:
blended_image = show_masks(show_boxes(image_np, boxes_filt), mask)
mask_images.append(Image.fromarray(blended_image))
image_np_copy = copy.deepcopy(image_np)
image_np_copy[~np.any(mask, axis=0)] = np.array([0, 0, 0, 0])
matted_images.append(Image.fromarray(image_np_copy))
return mask_images + masks_gallery + matted_images, sam_predict_status + sam_predict_result
def dino_predict(input_image, dino_model_name, text_prompt, box_threshold):
image_np = np.array(input_image)
boxes_filt, install_success = dino_predict_internal(input_image, dino_model_name, text_prompt, box_threshold)
if not install_success:
return None, gr.update(), gr.update(visible=True, value="GroundingDINO installment failed. Preview failed. See your terminal for more detail and submit an issue to https://github.com/continue-revolution/sd-webui-segment-anything/issues.")
boxes_filt = boxes_filt.numpy()
boxes_choice = [str(i) for i in range(boxes_filt.shape[0])]
return Image.fromarray(show_boxes(image_np, boxes_filt.astype(int), show_index=True)), gr.update(choices=boxes_choice, value=boxes_choice), gr.update(visible=False)
def dino_batch_process(
batch_sam_model_name, batch_dino_model_name, batch_text_prompt, batch_box_threshold, batch_dilation_amt,
dino_batch_source_dir, dino_batch_dest_dir,
dino_batch_output_per_image, dino_batch_save_mask, dino_batch_save_image_with_mask):
if batch_text_prompt is None or batch_text_prompt == "":
return "Please add text prompts to generate masks"
print("Start batch processing")
sam = init_sam_model(batch_sam_model_name)
predictor = SamPredictor(sam)
all_files = glob.glob(os.path.join(dino_batch_source_dir, "*"))
for image_index, input_image_file in enumerate(all_files):
print(f"processing {image_index}/{len(all_files)} {input_image_file}")
input_image = Image.open(input_image_file).convert("RGBA")
image_np = np.array(input_image)
image_np_rgb = image_np[..., :3]
boxes_filt, install_success = dino_predict_internal(input_image, batch_dino_model_name, batch_text_prompt, batch_box_threshold)
if not install_success:
return "GroundingDINO installment failed. Batch processing failed. See your terminal for more detail and submit an issue to https://github.com/continue-revolution/sd-webui-segment-anything/issues."
predictor.set_image(image_np_rgb)
transformed_boxes = predictor.transform.apply_boxes_torch(boxes_filt, image_np.shape[:2])
masks, _, _ = predictor.predict_torch(
point_coords=None,
point_labels=None,
boxes=transformed_boxes.to(device),
multimask_output=(dino_batch_output_per_image == 1),
)
masks = masks.permute(1, 0, 2, 3).cpu().numpy()
boxes_filt = boxes_filt.cpu().numpy().astype(int)
filename, ext = os.path.splitext(os.path.basename(input_image_file))
for idx, mask in enumerate(masks):
blended_image = show_masks(show_boxes(image_np, boxes_filt), mask)
merged_mask = np.any(mask, axis=0)
if batch_dilation_amt:
_, merged_mask = dilate_mask(merged_mask, batch_dilation_amt)
image_np_copy = copy.deepcopy(image_np)
image_np_copy[~merged_mask] = np.array([0, 0, 0, 0])
output_image = Image.fromarray(image_np_copy)
output_image.save(os.path.join(dino_batch_dest_dir, f"{filename}_{idx}_output{ext}"))
if dino_batch_save_mask:
output_mask = Image.fromarray(merged_mask)
output_mask.save(os.path.join(dino_batch_dest_dir, f"{filename}_{idx}_mask{ext}"))
if dino_batch_save_image_with_mask:
output_blend = Image.fromarray(blended_image)
output_blend.save(os.path.join(dino_batch_dest_dir, f"{filename}_{idx}_blend{ext}"))
if shared.cmd_opts.lowvram:
sam.to(cpu)
gc.collect()
torch_gc()
return "Done"
class Script(scripts.Script):
def title(self):
return 'Segment Anything'
def show(self, is_img2img):
return scripts.AlwaysVisible
def ui(self, is_img2img):
with gr.Accordion('Segment Anything', open=False, elem_id=id('accordion'), visible=is_img2img):
with gr.Tabs():
with gr.TabItem(label="Single Image"):
with gr.Column():
gr.HTML(value="<p>Left click the image to add one positive point (black dot). Right click the image to add one negative point (red dot). Left click the point to remove it.</p>")
with gr.Row():
sam_model_name = gr.Dropdown(label="SAM Model", elem_id="sam_model", choices=sam_model_list,
value=sam_model_list[0] if len(sam_model_list) > 0 else None)
refresh_models = ToolButton(value=refresh_symbol)
refresh_models.click(refresh_sam_models, sam_model_name, sam_model_name)
input_image = gr.Image(label="Image for Segment Anything", elem_id="sam_input_image",
show_label=False, source="upload", type="pil", image_mode="RGBA")
remove_dots = gr.Button(value="Remove all point prompts")
dummy_component = gr.Label(visible=False)
gr.HTML(value="<p>GroundingDINO + Segment Anything can achieve [text prompt]->[object detection]->[segmentation]</p>")
dino_checkbox = gr.Checkbox(value=False, label="Enable GroundingDINO", elem_id="dino_enable_checkbox")
with gr.Column(visible=False) as dino_column:
gr.HTML(value="<p>Due to the limitation of Segment Anything, when there are point prompts, at most 1 box prompt will be allowed; when there are multiple box prompts, no point prompts are allowed.</p>")
dino_model_name = gr.Dropdown(label="GroundingDINO Model (Auto download from huggingface)",
elem_id="dino_model", choices=dino_model_list, value=dino_model_list[0])
text_prompt = gr.Textbox(label="GroundingDINO Detection Prompt", elem_id="dino_text_prompt")
text_prompt.change(fn=lambda _: None, inputs=[dummy_component], outputs=None, _js="registerDinoTextObserver")
box_threshold = gr.Slider(label="GroundingDINO Box Threshold", minimum=0.0, maximum=1.0, value=0.3, step=0.001)
dino_preview_checkbox = gr.Checkbox(value=False, label="I want to preview GroundingDINO detection result and select the boxes I want.", elem_id="dino_preview_checkbox")
with gr.Column(visible=False) as dino_preview:
dino_preview_boxes = gr.Image(label="Image for GroundingDINO", elem_id="dino_box_output",
show_label=False, type="pil", image_mode="RGBA")
with gr.Row(elem_id="dino_generate_box", elem_classes="generate-box"):
gr.Button(value="Add text prompt to generate bounding box", elem_id="dino_no_button")
dino_preview_boxes_button = gr.Button(value="Generate bounding box", elem_id="dino_run_button")
dino_preview_boxes_selection = gr.CheckboxGroup(label="Select your favorite boxes: ", elem_id="dino_preview_boxes_selection")
dino_preview_boxes_selection.change(fn=lambda _: None, inputs=[dino_preview_boxes_selection], outputs=None, _js="onChangeDinoPreviewBoxesSelection")
dino_preview_result = gr.Text(value="", show_label=False, visible=False)
dino_preview_boxes_button.click(
fn=dino_predict,
_js="submit_dino",
inputs=[input_image, dino_model_name, text_prompt, box_threshold],
outputs=[dino_preview_boxes, dino_preview_boxes_selection, dino_preview_result])
mask_image = gr.Gallery(label='Segment Anything Output', show_label=False, elem_id='sam_gallery').style(grid=3)
with gr.Row(elem_id="sam_generate_box", elem_classes="generate-box"):
gr.Button(value="Add dot prompt or enable GroundingDINO with text prompts to preview segmentation", elem_id="sam_no_button")
run_button = gr.Button(value="Preview Segmentation", elem_id="sam_run_button")
run_result = gr.Text(value="", show_label=False)
gr.Checkbox(value=False, label="Preview automatically when add/remove points", elem_id="sam_realtime_preview_checkbox")
with gr.Row():
enabled = gr.Checkbox(value=False, label="Copy to Inpaint Upload", elem_id="sam_impaint_checkbox")
chosen_mask = gr.Radio(label="Choose your favorite mask: ", value="0", choices=["0", "1", "2"], type="index")
dilation_checkbox = gr.Checkbox(value=False, label="Expand Mask", elem_id="dilation_enable_checkbox")
with gr.Column(visible=False) as dilation_column:
dilation_amt = gr.Slider(minimum=0, maximum=100, default=0, value=0, label="Specify the amount that you wish to expand the mask by (recommend 30)", elem_id="dilation_amt")
expanded_mask_image = gr.Gallery(label="Expanded Mask", elem_id="sam_expanded_mask").style(grid=3)
update_mask_button = gr.Button(value="Update Mask", elem_id="update_mask_button")
switch = gr.Button(value="Switch to Inpaint Upload")
with gr.TabItem(label="Batch Process"):
gr.HTML(value="<p>You may configurate the following items and generate masked image for all images under a directory. This mode is designed for generating LoRA/LyCORIS training set.</p>")
gr.HTML(value="<p>The current workflow is [text prompt]->[object detection]->[segmentation]. Semantic segmentation support is coming soon!</p>")
with gr.Row():
batch_sam_model_name = gr.Dropdown(label="SAM Model", elem_id="batch_sam_model", choices=sam_model_list,
value=sam_model_list[0] if len(sam_model_list) > 0 else None)
batch_refresh_models = ToolButton(value=refresh_symbol)
batch_refresh_models.click(refresh_sam_models, sam_model_name, sam_model_name)
batch_dino_model_name = gr.Dropdown(label="GroundingDINO Model (Auto download from huggingface)",
elem_id="batch_dino_model", choices=dino_model_list, value=dino_model_list[0])
batch_text_prompt = gr.Textbox(label="GroundingDINO Detection Prompt", elem_id="batch_dino_text_prompt")
batch_box_threshold = gr.Slider(label="GroundingDINO Box Threshold", minimum=0.0, maximum=1.0, value=0.3, step=0.001)
batch_dilation_amt = gr.Slider(minimum=0, maximum=100, default=0, value=0, label="Specify the amount that you wish to expand the mask by (recommend 0-10)", elem_id="batch_dino_dilation_amt")
dino_batch_source_dir = gr.Textbox(label="Source directory")
dino_batch_dest_dir = gr.Textbox(label="Destination directory")
with gr.Row():
dino_batch_output_per_image = gr.Radio(
choices=["1", "3"], value="3", type="index", label="Output per image: ")
dino_batch_save_mask = gr.Checkbox(value=True, label="Save mask")
dino_batch_save_image_with_mask = gr.Checkbox(
value=True, label="Save original image with mask and bounding box")
dino_batch_run_button = gr.Button(value="Start batch process")
dino_batch_progress = gr.Text(value="", show_label=False)
dino_batch_run_button.click(
fn=dino_batch_process,
inputs=[batch_sam_model_name, batch_dino_model_name, batch_text_prompt, batch_box_threshold, batch_dilation_amt,
dino_batch_source_dir, dino_batch_dest_dir,
dino_batch_output_per_image, dino_batch_save_mask,
dino_batch_save_image_with_mask],
outputs=[dino_batch_progress]
)
unload = gr.Button(value="Unload all models from memory")
run_button.click(
fn=sam_predict,
_js='submit_sam',
inputs=[sam_model_name, input_image, # SAM
dummy_component, dummy_component, # Point prompts
dino_checkbox, dino_model_name, text_prompt, box_threshold, # DINO prompts
dino_preview_checkbox, dino_preview_boxes_selection], # DINO preview prompts
outputs=[mask_image, run_result])
dino_checkbox.change(
fn=gr_show,
inputs=[dino_checkbox],
outputs=[dino_column],
show_progress=False)
dino_preview_checkbox.change(
fn=gr_show,
inputs=[dino_preview_checkbox],
outputs=[dino_preview],
show_progress=False)
switch.click(
fn=lambda _: None,
_js="switchToInpaintUpload",
inputs=[dummy_component],
outputs=None)
remove_dots.click(
fn=lambda _: None,
_js="removeDots",
inputs=[dummy_component],
outputs=None)
unload.click(
fn=clear_cache,
inputs=[],
outputs=[])
dilation_checkbox.change(
fn=gr_show,
inputs=[dilation_checkbox],
outputs=[dilation_column],
show_progress=False)
update_mask_button.click(
fn=update_mask,
inputs=[mask_image, chosen_mask, dilation_amt, input_image],
outputs=[expanded_mask_image])
return [enabled, input_image, mask_image, chosen_mask, dilation_checkbox, expanded_mask_image]
def process(self, p: StableDiffusionProcessingImg2Img, enabled=False, input_image=None, mask=None, chosen_mask=0, dilation_enabled=False, expanded_mask=None):
if not enabled or input_image is None or mask is None or not isinstance(p, StableDiffusionProcessingImg2Img):
return
p.init_images = [input_image]
p.image_mask = Image.open(expanded_mask[1]['name'] if dilation_enabled and expanded_mask is not None else mask[chosen_mask + 3]['name'])