feat: extract safety checker, remove dependency to diffusers

extras/diffusers/pipelines/stable_diffusion/safety_checker.py

@@ -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

modules/censor.py

@@ -1,10 +1,7 @@
 # modified version of https://github.com/AUTOMATIC1111/stable-diffusion-webui-nsfw-censor/blob/master/scripts/censor.py
-
 import numpy as np
-import torch
-import modules.core as core
 
-from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from extras.diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from transformers import AutoFeatureExtractor
 from PIL import Image
 import modules.config
@@ -16,8 +13,6 @@ safety_checker = None
 
 def numpy_to_pil(image):
     image = (image * 255).round().astype("uint8")
-
-    #pil_image = Image.fromarray(image, 'RGB')
     pil_image = Image.fromarray(image)
 
     return pil_image
@@ -27,7 +22,7 @@ def numpy_to_pil(image):
 def check_safety(x_image):
     global safety_feature_extractor, safety_checker
 
-    if safety_feature_extractor is None:
+    if safety_feature_extractor is None or safety_checker is None:
         safety_feature_extractor = AutoFeatureExtractor.from_pretrained(safety_model_id, cache_dir=modules.config.path_safety_checker_models)
         safety_checker = StableDiffusionSafetyChecker.from_pretrained(safety_model_id, cache_dir=modules.config.path_safety_checker_models)
 
@@ -52,4 +47,4 @@ def censor_single(x):
 def censor_batch(images):
     images = [censor_single(image) for image in images]
 
-    return images
+    return images

requirements_versions.txt

@@ -16,4 +16,3 @@ opencv-contrib-python==4.8.0.74
 httpx==0.24.1
 onnxruntime==1.16.3
 timm==0.9.2
-diffusers==0.25.1