happyz
/
Fooocus
mirror of https://github.com/lllyasviel/Fooocus.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

feat: add nsfw image censoring via config and checkbox (#958) 

* add nsfw image censoring

activatable via config, uses CompVis/stable-diffusion-safety-checker

* fix progressbar call for nsfw output

* use config to set cache dir for safety checker

* add checkbox black_out_nsfw

makes both enabling via config and checkbox possible, where config overrides the checkbox value

* fix: add missing diffusers package

* feat: extract safety checker, remove dependency to diffusers

* feat: make code compatible again after merge with main

* feat: move censor to extras, optimize safety checker file handling

* refactor: rename folder safety_checker_models to safety_checker
This commit is contained in:
Manuel Schmid 2024-05-18 15:50:28 +02:00 committed by GitHub
parent 33fa175bd4
commit 00d3d1b4b3
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
9 changed files with 424 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

56
extras/censor.py Normal file
View File

@ -0,0 +1,56 @@
# modified version of https://github.com/AUTOMATIC1111/stable-diffusion-webui-nsfw-censor/blob/master/scripts/censor.py
import numpy as np
import os
from extras.safety_checker.models.safety_checker import StableDiffusionSafetyChecker
from transformers import CLIPFeatureExtractor, CLIPConfig
from PIL import Image
import modules.config
safety_checker_repo_root = os.path.join(os.path.dirname(__file__), 'safety_checker')
config_path = os.path.join(safety_checker_repo_root, "configs", "config.json")
preprocessor_config_path = os.path.join(safety_checker_repo_root, "configs", "preprocessor_config.json")
safety_feature_extractor = None
safety_checker = None
def numpy_to_pil(image):
image = (image * 255).round().astype("uint8")
pil_image = Image.fromarray(image)
return pil_image
# check and replace nsfw content
def check_safety(x_image):
global safety_feature_extractor, safety_checker
if safety_feature_extractor is None or safety_checker is None:
safety_checker_model = modules.config.downloading_safety_checker_model()
safety_feature_extractor = CLIPFeatureExtractor.from_json_file(preprocessor_config_path)
clip_config = CLIPConfig.from_json_file(config_path)
safety_checker = StableDiffusionSafetyChecker.from_pretrained(safety_checker_model, config=clip_config)
safety_checker_input = safety_feature_extractor(numpy_to_pil(x_image), return_tensors="pt")
x_checked_image, has_nsfw_concept = safety_checker(images=x_image, clip_input=safety_checker_input.pixel_values)
return x_checked_image, has_nsfw_concept
def censor_single(x):
x_checked_image, has_nsfw_concept = check_safety(x)
# replace image with black pixels, keep dimensions
# workaround due to different numpy / pytorch image matrix format
if has_nsfw_concept[0]:
imageshape = x_checked_image.shape
x_checked_image = np.zeros((imageshape[0], imageshape[1], 3), dtype = np.uint8)
return x_checked_image
def censor_batch(images):
images = [censor_single(image) for image in images]
return images

171
extras/safety_checker/configs/config.json Normal file
View File

@ -0,0 +1,171 @@
{
"_name_or_path": "clip-vit-large-patch14/",
"architectures": [
"SafetyChecker"
],
"initializer_factor": 1.0,
"logit_scale_init_value": 2.6592,
"model_type": "clip",
"projection_dim": 768,
"text_config": {
"_name_or_path": "",
"add_cross_attention": false,
"architectures": null,
"attention_dropout": 0.0,
"bad_words_ids": null,
"bos_token_id": 0,
"chunk_size_feed_forward": 0,
"cross_attention_hidden_size": null,
"decoder_start_token_id": null,
"diversity_penalty": 0.0,
"do_sample": false,
"dropout": 0.0,
"early_stopping": false,
"encoder_no_repeat_ngram_size": 0,
"eos_token_id": 2,
"exponential_decay_length_penalty": null,
"finetuning_task": null,
"forced_bos_token_id": null,
"forced_eos_token_id": null,
"hidden_act": "quick_gelu",
"hidden_size": 768,
"id2label": {
"0": "LABEL_0",
"1": "LABEL_1"
},
"initializer_factor": 1.0,
"initializer_range": 0.02,
"intermediate_size": 3072,
"is_decoder": false,
"is_encoder_decoder": false,
"label2id": {
"LABEL_0": 0,
"LABEL_1": 1
},
"layer_norm_eps": 1e-05,
"length_penalty": 1.0,
"max_length": 20,
"max_position_embeddings": 77,
"min_length": 0,
"model_type": "clip_text_model",
"no_repeat_ngram_size": 0,
"num_attention_heads": 12,
"num_beam_groups": 1,
"num_beams": 1,
"num_hidden_layers": 12,
"num_return_sequences": 1,
"output_attentions": false,
"output_hidden_states": false,
"output_scores": false,
"pad_token_id": 1,
"prefix": null,
"problem_type": null,
"pruned_heads": {},
"remove_invalid_values": false,
"repetition_penalty": 1.0,
"return_dict": true,
"return_dict_in_generate": false,
"sep_token_id": null,
"task_specific_params": null,
"temperature": 1.0,
"tie_encoder_decoder": false,
"tie_word_embeddings": true,
"tokenizer_class": null,
"top_k": 50,
"top_p": 1.0,
"torch_dtype": null,
"torchscript": false,
"transformers_version": "4.21.0.dev0",
"typical_p": 1.0,
"use_bfloat16": false,
"vocab_size": 49408
},
"text_config_dict": {
"hidden_size": 768,
"intermediate_size": 3072,
"num_attention_heads": 12,
"num_hidden_layers": 12
},
"torch_dtype": "float32",
"transformers_version": null,
"vision_config": {
"_name_or_path": "",
"add_cross_attention": false,
"architectures": null,
"attention_dropout": 0.0,
"bad_words_ids": null,
"bos_token_id": null,
"chunk_size_feed_forward": 0,
"cross_attention_hidden_size": null,
"decoder_start_token_id": null,
"diversity_penalty": 0.0,
"do_sample": false,
"dropout": 0.0,
"early_stopping": false,
"encoder_no_repeat_ngram_size": 0,
"eos_token_id": null,
"exponential_decay_length_penalty": null,
"finetuning_task": null,
"forced_bos_token_id": null,
"forced_eos_token_id": null,
"hidden_act": "quick_gelu",
"hidden_size": 1024,
"id2label": {
"0": "LABEL_0",
"1": "LABEL_1"
},
"image_size": 224,
"initializer_factor": 1.0,
"initializer_range": 0.02,
"intermediate_size": 4096,
"is_decoder": false,
"is_encoder_decoder": false,
"label2id": {
"LABEL_0": 0,
"LABEL_1": 1
},
"layer_norm_eps": 1e-05,
"length_penalty": 1.0,
"max_length": 20,
"min_length": 0,
"model_type": "clip_vision_model",
"no_repeat_ngram_size": 0,
"num_attention_heads": 16,
"num_beam_groups": 1,
"num_beams": 1,
"num_hidden_layers": 24,
"num_return_sequences": 1,
"output_attentions": false,
"output_hidden_states": false,
"output_scores": false,
"pad_token_id": null,
"patch_size": 14,
"prefix": null,
"problem_type": null,
"pruned_heads": {},
"remove_invalid_values": false,
"repetition_penalty": 1.0,
"return_dict": true,
"return_dict_in_generate": false,
"sep_token_id": null,
"task_specific_params": null,
"temperature": 1.0,
"tie_encoder_decoder": false,
"tie_word_embeddings": true,
"tokenizer_class": null,
"top_k": 50,
"top_p": 1.0,
"torch_dtype": null,
"torchscript": false,
"transformers_version": "4.21.0.dev0",
"typical_p": 1.0,
"use_bfloat16": false
},
"vision_config_dict": {
"hidden_size": 1024,
"intermediate_size": 4096,
"num_attention_heads": 16,
"num_hidden_layers": 24,
"patch_size": 14
}
}

20
extras/safety_checker/configs/preprocessor_config.json Normal file
View File

@ -0,0 +1,20 @@
{
"crop_size": 224,
"do_center_crop": true,
"do_convert_rgb": true,
"do_normalize": true,
"do_resize": true,
"feature_extractor_type": "CLIPFeatureExtractor",
"image_mean": [
0.48145466,
0.4578275,
0.40821073
],
"image_std": [
0.26862954,
0.26130258,
0.27577711
],
"resample": 3,
"size": 224
}

126
extras/safety_checker/models/safety_checker.py Normal file
View File

@ -0,0 +1,126 @@
# from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/safety_checker.py
# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import torch
import torch.nn as nn
from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel
from transformers.utils import logging
logger = logging.get_logger(__name__)
def cosine_distance(image_embeds, text_embeds):
normalized_image_embeds = nn.functional.normalize(image_embeds)
normalized_text_embeds = nn.functional.normalize(text_embeds)
return torch.mm(normalized_image_embeds, normalized_text_embeds.t())
class StableDiffusionSafetyChecker(PreTrainedModel):
config_class = CLIPConfig
main_input_name = "clip_input"
_no_split_modules = ["CLIPEncoderLayer"]
def __init__(self, config: CLIPConfig):
super().__init__(config)
self.vision_model = CLIPVisionModel(config.vision_config)
self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False)
self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False)
self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False)
self.concept_embeds_weights = nn.Parameter(torch.ones(17), requires_grad=False)
self.special_care_embeds_weights = nn.Parameter(torch.ones(3), requires_grad=False)
@torch.no_grad()
def forward(self, clip_input, images):
pooled_output = self.vision_model(clip_input)[1] # pooled_output
image_embeds = self.visual_projection(pooled_output)
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().float().numpy()
cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().float().numpy()
result = []
batch_size = image_embeds.shape[0]
for i in range(batch_size):
result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []}
# increase this value to create a stronger `nfsw` filter
# at the cost of increasing the possibility of filtering benign images
adjustment = 0.0
for concept_idx in range(len(special_cos_dist[0])):
concept_cos = special_cos_dist[i][concept_idx]
concept_threshold = self.special_care_embeds_weights[concept_idx].item()
result_img["special_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3)
if result_img["special_scores"][concept_idx] > 0:
result_img["special_care"].append({concept_idx, result_img["special_scores"][concept_idx]})
adjustment = 0.01
for concept_idx in range(len(cos_dist[0])):
concept_cos = cos_dist[i][concept_idx]
concept_threshold = self.concept_embeds_weights[concept_idx].item()
result_img["concept_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3)
if result_img["concept_scores"][concept_idx] > 0:
result_img["bad_concepts"].append(concept_idx)
result.append(result_img)
has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result]
for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
if has_nsfw_concept:
if torch.is_tensor(images) or torch.is_tensor(images[0]):
images[idx] = torch.zeros_like(images[idx]) # black image
else:
images[idx] = np.zeros(images[idx].shape) # black image
if any(has_nsfw_concepts):
logger.warning(
"Potential NSFW content was detected in one or more images. A black image will be returned instead."
" Try again with a different prompt and/or seed."
)
return images, has_nsfw_concepts
@torch.no_grad()
def forward_onnx(self, clip_input: torch.Tensor, images: torch.Tensor):
pooled_output = self.vision_model(clip_input)[1] # pooled_output
image_embeds = self.visual_projection(pooled_output)
special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds)
cos_dist = cosine_distance(image_embeds, self.concept_embeds)
# increase this value to create a stronger `nsfw` filter
# at the cost of increasing the possibility of filtering benign images
adjustment = 0.0
special_scores = special_cos_dist - self.special_care_embeds_weights + adjustment
# special_scores = special_scores.round(decimals=3)
special_care = torch.any(special_scores > 0, dim=1)
special_adjustment = special_care * 0.01
special_adjustment = special_adjustment.unsqueeze(1).expand(-1, cos_dist.shape[1])
concept_scores = (cos_dist - self.concept_embeds_weights) + special_adjustment
# concept_scores = concept_scores.round(decimals=3)
has_nsfw_concepts = torch.any(concept_scores > 0, dim=1)
images[has_nsfw_concepts] = 0.0 # black image
return images, has_nsfw_concepts

2
language/en.json
View File

@ -55,6 +55,8 @@
"Disable seed increment": "Disable seed increment",
"Disable automatic seed increment when image number is > 1.": "Disable automatic seed increment when image number is > 1.",
"Read wildcards in order": "Read wildcards in order",
"Black Out NSFW": "Black Out NSFW",
"Use black image if NSFW is detected.": "Use black image if NSFW is detected.",
"\ud83d\udcda History Log": "\uD83D\uDCDA History Log",
"Image Style": "Image Style",
"Fooocus V2": "Fooocus V2",

0
models/safety_checker/put_safety_checker_models_here Normal file
View File

32
modules/async_worker.py
View File

@ -43,6 +43,7 @@ def worker():
import fooocus_version
import args_manager
from extras.censor import censor_batch, censor_single
from modules.sdxl_styles import apply_style, get_random_style, apply_wildcards, fooocus_expansion, apply_arrays, random_style_name
from modules.private_logger import log
from extras.expansion import safe_str
@ -68,10 +69,14 @@ def worker():
print(f'[Fooocus] {text}')
async_task.yields.append(['preview', (number, text, None)])
def yield_result(async_task, imgs, do_not_show_finished_images=False):
def yield_result(async_task, imgs, black_out_nsfw, censor=True, do_not_show_finished_images=False, progressbar_index=13):
if not isinstance(imgs, list):
imgs = [imgs]
if censor and (modules.config.default_black_out_nsfw or black_out_nsfw):
progressbar(async_task, progressbar_index, 'Checking for NSFW content ...')
imgs = censor_batch(imgs)
async_task.results = async_task.results + imgs
if do_not_show_finished_images:
@ -160,6 +165,7 @@ def worker():
disable_preview = args.pop()
disable_intermediate_results = args.pop()
disable_seed_increment = args.pop()
black_out_nsfw = args.pop()
adm_scaler_positive = args.pop()
adm_scaler_negative = args.pop()
adm_scaler_end = args.pop()
@ -578,8 +584,11 @@ def worker():
if direct_return:
d = [('Upscale (Fast)', 'upscale_fast', '2x')]
if modules.config.default_black_out_nsfw or black_out_nsfw:
progressbar(async_task, 100, 'Checking for NSFW content ...')
uov_input_image = censor_single(uov_input_image)
uov_input_image_path = log(uov_input_image, d, output_format=output_format)
yield_result(async_task, uov_input_image_path, do_not_show_finished_images=True)
yield_result(async_task, uov_input_image_path, black_out_nsfw, False, do_not_show_finished_images=True)
return
tiled = True
@ -643,8 +652,7 @@ def worker():
)
if debugging_inpaint_preprocessor:
yield_result(async_task, inpaint_worker.current_task.visualize_mask_processing(),
do_not_show_finished_images=True)
yield_result(async_task, inpaint_worker.current_task.visualize_mask_processing(), black_out_nsfw, do_not_show_finished_images=True)
return
progressbar(async_task, 13, 'VAE Inpaint encoding ...')
@ -707,7 +715,7 @@ def worker():
cn_img = HWC3(cn_img)
task[0] = core.numpy_to_pytorch(cn_img)
if debugging_cn_preprocessor:
yield_result(async_task, cn_img, do_not_show_finished_images=True)
yield_result(async_task, cn_img, black_out_nsfw, do_not_show_finished_images=True)
return
for task in cn_tasks[flags.cn_cpds]:
cn_img, cn_stop, cn_weight = task
@ -719,7 +727,7 @@ def worker():
cn_img = HWC3(cn_img)
task[0] = core.numpy_to_pytorch(cn_img)
if debugging_cn_preprocessor:
yield_result(async_task, cn_img, do_not_show_finished_images=True)
yield_result(async_task, cn_img, black_out_nsfw, do_not_show_finished_images=True)
return
for task in cn_tasks[flags.cn_ip]:
cn_img, cn_stop, cn_weight = task
@ -730,7 +738,7 @@ def worker():
task[0] = ip_adapter.preprocess(cn_img, ip_adapter_path=ip_adapter_path)
if debugging_cn_preprocessor:
yield_result(async_task, cn_img, do_not_show_finished_images=True)
yield_result(async_task, cn_img, black_out_nsfw, do_not_show_finished_images=True)
return
for task in cn_tasks[flags.cn_ip_face]:
cn_img, cn_stop, cn_weight = task
@ -744,7 +752,7 @@ def worker():
task[0] = ip_adapter.preprocess(cn_img, ip_adapter_path=ip_adapter_face_path)
if debugging_cn_preprocessor:
yield_result(async_task, cn_img, do_not_show_finished_images=True)
yield_result(async_task, cn_img, black_out_nsfw, do_not_show_finished_images=True)
return
all_ip_tasks = cn_tasks[flags.cn_ip] + cn_tasks[flags.cn_ip_face]
@ -844,6 +852,12 @@ def worker():
imgs = [inpaint_worker.current_task.post_process(x) for x in imgs]
img_paths = []
if modules.config.default_black_out_nsfw or black_out_nsfw:
progressbar(async_task, int(15.0 + 85.0 * float((current_task_id + 1) * steps) / float(all_steps)),
'Checking for NSFW content ...')
imgs = censor_batch(imgs)
for x in imgs:
d = [('Prompt', 'prompt', task['log_positive_prompt']),
('Negative Prompt', 'negative_prompt', task['log_negative_prompt']),
@ -895,7 +909,7 @@ def worker():
d.append(('Version', 'version', 'Fooocus v' + fooocus_version.version))
img_paths.append(log(x, d, metadata_parser, output_format, task))
yield_result(async_task, img_paths, do_not_show_finished_images=len(tasks) == 1 or disable_intermediate_results)
yield_result(async_task, img_paths, black_out_nsfw, False, do_not_show_finished_images=len(tasks) == 1 or disable_intermediate_results)
except ldm_patched.modules.model_management.InterruptProcessingException as e:
if async_task.last_stop == 'skip':
print('User skipped')

14
modules/config.py
View File

@ -196,6 +196,7 @@ path_controlnet = get_dir_or_set_default('path_controlnet', '../models/controlne
path_clip_vision = get_dir_or_set_default('path_clip_vision', '../models/clip_vision/')
path_fooocus_expansion = get_dir_or_set_default('path_fooocus_expansion', '../models/prompt_expansion/fooocus_expansion')
path_wildcards = get_dir_or_set_default('path_wildcards', '../wildcards/')
path_safety_checker = get_dir_or_set_default('path_safety_checker', '../models/safety_checker/')
path_outputs = get_path_output()
@ -456,6 +457,11 @@ example_inpaint_prompts = get_config_item_or_set_default(
],
validator=lambda x: isinstance(x, list) and all(isinstance(v, str) for v in x)
)
default_black_out_nsfw = get_config_item_or_set_default(
key='default_black_out_nsfw',
default_value=False,
validator=lambda x: isinstance(x, bool)
)
default_save_metadata_to_images = get_config_item_or_set_default(
key='default_save_metadata_to_images',
default_value=False,
@ -691,5 +697,13 @@ def downloading_upscale_model():
)
return os.path.join(path_upscale_models, 'fooocus_upscaler_s409985e5.bin')
def downloading_safety_checker_model():
load_file_from_url(
url='https://huggingface.co/mashb1t/misc/resolve/main/stable-diffusion-safety-checker.bin',
model_dir=path_safety_checker,
file_name='stable-diffusion-safety-checker.bin'
)
return os.path.join(path_safety_checker, 'stable-diffusion-safety-checker.bin')
update_files()

14
webui.py
View File

@ -436,7 +436,8 @@ with shared.gradio_root:
overwrite_upscale_strength = gr.Slider(label='Forced Overwrite of Denoising Strength of "Upscale"',
minimum=-1, maximum=1.0, step=0.001, value=-1,
info='Set as negative number to disable. For developer debugging.')
disable_preview = gr.Checkbox(label='Disable Preview', value=False,
disable_preview = gr.Checkbox(label='Disable Preview', value=modules.config.default_black_out_nsfw,
interactive=not modules.config.default_black_out_nsfw,
info='Disable preview during generation.')
disable_intermediate_results = gr.Checkbox(label='Disable Intermediate Results',
value=modules.config.default_performance == flags.Performance.EXTREME_SPEED.value,
@ -447,6 +448,15 @@ with shared.gradio_root:
value=False)
read_wildcards_in_order = gr.Checkbox(label="Read wildcards in order", value=False)
black_out_nsfw = gr.Checkbox(label='Black Out NSFW',
value=modules.config.default_black_out_nsfw,
interactive=not modules.config.default_black_out_nsfw,
info='Use black image if NSFW is detected.')
black_out_nsfw.change(lambda x: gr.update(value=x, interactive=not x),
inputs=black_out_nsfw, outputs=disable_preview, queue=False,
show_progress=False)
if not args_manager.args.disable_metadata:
save_metadata_to_images = gr.Checkbox(label='Save Metadata to Images', value=modules.config.default_save_metadata_to_images,
info='Adds parameters to generated images allowing manual regeneration.')
@ -636,7 +646,7 @@ with shared.gradio_root:
ctrls += [input_image_checkbox, current_tab]
ctrls += [uov_method, uov_input_image]
ctrls += [outpaint_selections, inpaint_input_image, inpaint_additional_prompt, inpaint_mask_image]
ctrls += [disable_preview, disable_intermediate_results, disable_seed_increment]
ctrls += [disable_preview, disable_intermediate_results, disable_seed_increment, black_out_nsfw]
ctrls += [adm_scaler_positive, adm_scaler_negative, adm_scaler_end, adaptive_cfg]
ctrls += [sampler_name, scheduler_name, vae_name]
ctrls += [overwrite_step, overwrite_switch, overwrite_width, overwrite_height, overwrite_vary_strength]