feature: improved safety filter

- provides more informative logs
- provides a detailed safety score object
- adds non-bypassable filter for extreme content

README.md

@@ -213,6 +213,8 @@ docker run -it --gpus all -v $HOME/.cache/huggingface:/root/.cache/huggingface -
 [Example Colab](https://colab.research.google.com/drive/1rOvQNs0Cmn_yU1bKWjCOHzGVDgZkaTtO?usp=sharing)
 
 ## ChangeLog
+**3.0.0**
+- feature: improved safety filter
 
 **2.4.0**
 - 🎉 feature: prompt expansion

imaginairy/api.py

@@ -24,7 +24,7 @@ from imaginairy.img_log import (
     log_latent,
 )
 from imaginairy.img_utils import pillow_fit_image_within, pillow_img_to_torch_image
-from imaginairy.safety import is_nsfw
+from imaginairy.safety import SafetyMode, create_safety_score
 from imaginairy.samplers.base import get_sampler
 from imaginairy.samplers.plms import PLMSSchedule
 from imaginairy.schema import ImaginePrompt, ImagineResult
@@ -40,16 +40,14 @@ LIB_PATH = os.path.dirname(__file__)
 logger = logging.getLogger(__name__)
 
 
-class SafetyMode:
-    DISABLED = "disabled"
-    CLASSIFY = "classify"
-    FILTER = "filter"
-
-
 # leave undocumented. I'd ask that no one publicize this flag. Just want a
 # slight barrier to entry. Please don't use this is any way that's gonna cause
-# the press or governments to freak out about AI...
-IMAGINAIRY_SAFETY_MODE = os.getenv("IMAGINAIRY_SAFETY_MODE", SafetyMode.FILTER)
+# the media or politicians to freak out about AI...
+IMAGINAIRY_SAFETY_MODE = os.getenv("IMAGINAIRY_SAFETY_MODE", SafetyMode.STRICT)
+if IMAGINAIRY_SAFETY_MODE in {"disabled", "classify"}:
+    IMAGINAIRY_SAFETY_MODE = SafetyMode.RELAXED
+elif IMAGINAIRY_SAFETY_MODE == "filter":
+    IMAGINAIRY_SAFETY_MODE = SafetyMode.STRICT
 
 DEFAULT_MODEL_WEIGHTS_LOCATION = (
     "https://www.googleapis.com/storage/v1/b/aai-blog-files/o/sd-v1-4.ckpt?alt=media"
@@ -68,8 +66,7 @@ def load_model_from_config(
         ckpt_path = cached_path(model_weights_location)
     else:
         ckpt_path = model_weights_location
-    logger.info(f"Loading model onto {get_device()} backend...")
-    logger.debug(f"Loading model from {ckpt_path}")
+    logger.info(f"Loading model {ckpt_path} onto {get_device()} backend...")
     pl_sd = torch.load(ckpt_path, map_location="cpu")
     if "global_step" in pl_sd:
         logger.debug(f"Global Step: {pl_sd['global_step']}")
@@ -352,59 +349,61 @@ def imagine(
 
                     upscaled_img = None
                     rebuilt_orig_img = None
-                    is_nsfw_img = None
+
                     if add_caption:
                         caption = generate_caption(img)
                         logger.info(f"    Generated caption: {caption}")
-                    if IMAGINAIRY_SAFETY_MODE != SafetyMode.DISABLED:
-                        is_nsfw_img = is_nsfw(img)
-                        if is_nsfw_img and IMAGINAIRY_SAFETY_MODE == SafetyMode.FILTER:
-                            logger.info("    ⚠️  Filtering NSFW image")
-                            img = img.filter(ImageFilter.GaussianBlur(radius=40))
-
-                    if prompt.fix_faces:
-                        logger.info("    Fixing 😊 's in 🖼  using CodeFormer...")
-                        img = enhance_faces(img, fidelity=prompt.fix_faces_fidelity)
-                    if prompt.upscale:
-                        logger.info("    Upscaling 🖼  using real-ESRGAN...")
-                        upscaled_img = upscale_image(img)
-
-                    # put the newly generated patch back into the original, full size image
-                    if (
-                        prompt.mask_modify_original
-                        and mask_image_orig
-                        and prompt.init_image
-                    ):
-                        img_to_add_back_to_original = (
-                            upscaled_img if upscaled_img else img
-                        )
-                        img_to_add_back_to_original = (
-                            img_to_add_back_to_original.resize(
+
+                    safety_score = create_safety_score(
+                        img,
+                        safety_mode=IMAGINAIRY_SAFETY_MODE,
+                    )
+                    if not safety_score.is_filtered:
+                        if prompt.fix_faces:
+                            logger.info("    Fixing 😊 's in 🖼  using CodeFormer...")
+                            img = enhance_faces(img, fidelity=prompt.fix_faces_fidelity)
+                        if prompt.upscale:
+                            logger.info("    Upscaling 🖼  using real-ESRGAN...")
+                            upscaled_img = upscale_image(img)
+
+                        # put the newly generated patch back into the original, full size image
+                        if (
+                            prompt.mask_modify_original
+                            and mask_image_orig
+                            and prompt.init_image
+                        ):
+                            img_to_add_back_to_original = (
+                                upscaled_img if upscaled_img else img
+                            )
+                            img_to_add_back_to_original = (
+                                img_to_add_back_to_original.resize(
+                                    prompt.init_image.size,
+                                    resample=Image.Resampling.LANCZOS,
+                                )
+                            )
+
+                            mask_for_orig_size = mask_image_orig.resize(
                                 prompt.init_image.size,
                                 resample=Image.Resampling.LANCZOS,
                             )
-                        )
-
-                        mask_for_orig_size = mask_image_orig.resize(
-                            prompt.init_image.size, resample=Image.Resampling.LANCZOS
-                        )
-                        mask_for_orig_size = mask_for_orig_size.filter(
-                            ImageFilter.GaussianBlur(radius=5)
-                        )
-                        log_img(mask_for_orig_size, "mask for original image size")
+                            mask_for_orig_size = mask_for_orig_size.filter(
+                                ImageFilter.GaussianBlur(radius=5)
+                            )
+                            log_img(mask_for_orig_size, "mask for original image size")
 
-                        rebuilt_orig_img = Image.composite(
-                            prompt.init_image,
-                            img_to_add_back_to_original,
-                            mask_for_orig_size,
-                        )
-                        log_img(rebuilt_orig_img, "reconstituted original")
+                            rebuilt_orig_img = Image.composite(
+                                prompt.init_image,
+                                img_to_add_back_to_original,
+                                mask_for_orig_size,
+                            )
+                            log_img(rebuilt_orig_img, "reconstituted original")
 
                     yield ImagineResult(
                         img=img,
                         prompt=prompt,
                         upscaled_img=upscaled_img,
-                        is_nsfw=is_nsfw_img,
+                        is_nsfw=safety_score.is_nsfw,
+                        safety_score=safety_score,
                         modified_original=rebuilt_orig_img,
                         mask_binary=mask_image_orig,
                         mask_grayscale=mask_grayscale,

imaginairy/safety.py

@@ -1,20 +1,128 @@
+import logging
 from functools import lru_cache
 
-import numpy as np
 import torch
 from diffusers.pipelines.stable_diffusion import safety_checker as safety_checker_mod
-from diffusers.pipelines.stable_diffusion.safety_checker import (
-    StableDiffusionSafetyChecker,
-)
 from transformers import AutoFeatureExtractor
 
+logger = logging.getLogger(__name__)
+
+
+class SafetyMode:
+    STRICT = "strict"
+    RELAXED = "relaxed"
+
+
+class SafetyResult:
+    # increase this value to create a stronger `nfsw` filter
+    # at the cost of increasing the possibility of filtering benign images
+    _default_adjustment = 0.0
+
+    def __init__(self):
+        self.nsfw_scores = {}
+        self.special_care_scores = {}
+        self.is_filtered = False
+
+    def add_special_care_score(self, concept_idx, abs_score, threshold):
+        adjustment = self._default_adjustment
+        adjusted_score = round(abs_score - threshold + adjustment, 3)
+        try:
+            score_name = _SPECIAL_CARE_DESCRIPTIONS[concept_idx]
+        except LookupError:
+            score_name = ""
+        if adjusted_score > 0 and score_name:
+            logger.debug(
+                f"    🔞🔞 '{score_name}' abs:{abs_score:.3f} adj:{adjusted_score}"
+            )
+        self.special_care_scores[concept_idx] = adjusted_score
+
+    def add_nsfw_score(self, concept_idx, abs_score, threshold):
+        if len(self.special_care_scores) != 3:
+            raise ValueError("special care scores must be set first")
+        adjustment = self._default_adjustment
+        if self.special_care_score > 0:
+            adjustment += 0.01
+        adjusted_score = round(abs_score - threshold + adjustment, 3)
+        try:
+            score_name = _CONCEPT_DESCRIPTIONS[concept_idx]
+        except LookupError:
+            score_name = ""
+        if adjusted_score > 0 and score_name:
+            logger.debug(
+                f"    🔞 '{concept_idx}:{score_name}' abs:{abs_score:.3f} adj:{adjusted_score}"
+            )
+        self.nsfw_scores[concept_idx] = adjusted_score
+
+    @property
+    def nsfw_score(self):
+        return max(self.nsfw_scores.values())
+
+    @property
+    def special_care_score(self):
+        return max(self.special_care_scores.values())
+
+    @property
+    def special_care_nsfw_score(self):
+        return min(self.nsfw_score, self.special_care_score)
+
+    @property
+    def is_nsfw(self):
+        return self.nsfw_score > 0
+
+    @property
+    def is_special_care_nsfw(self):
+        return self.special_care_nsfw_score > 0
+
+
+class EnhancedStableDiffusionSafetyChecker(
+    safety_checker_mod.StableDiffusionSafetyChecker
+):
+    @torch.no_grad()
+    def forward(self, clip_input):
+        pooled_output = self.vision_model(clip_input)[1]
+        image_embeds = self.visual_projection(pooled_output)
+
+        special_cos_dist = (
+            safety_checker_mod.cosine_distance(image_embeds, self.special_care_embeds)
+            .cpu()
+            .numpy()
+        )
+        cos_dist = (
+            safety_checker_mod.cosine_distance(image_embeds, self.concept_embeds)
+            .cpu()
+            .numpy()
+        )
+
+        safety_results = []
+        batch_size = image_embeds.shape[0]
+        for i in range(batch_size):
+            safety_result = SafetyResult()
+
+            for concet_idx in range(len(special_cos_dist[0])):
+                concept_cos = special_cos_dist[i][concet_idx]
+                concept_threshold = self.special_care_embeds_weights[concet_idx].item()
+                safety_result.add_special_care_score(
+                    concet_idx, concept_cos, concept_threshold
+                )
+
+            for concet_idx in range(len(cos_dist[0])):
+                concept_cos = cos_dist[i][concet_idx]
+                concept_threshold = self.concept_embeds_weights[concet_idx].item()
+                safety_result.add_nsfw_score(concet_idx, concept_cos, concept_threshold)
+
+            safety_results.append(safety_result)
+
+        return safety_results
+
 
 @lru_cache()
 def safety_models():
     safety_model_id = "CompVis/stable-diffusion-safety-checker"
     monkeypatch_safety_cosine_distance()
     safety_feature_extractor = AutoFeatureExtractor.from_pretrained(safety_model_id)
-    safety_checker = StableDiffusionSafetyChecker.from_pretrained(safety_model_id)
+    safety_checker = EnhancedStableDiffusionSafetyChecker.from_pretrained(
+        safety_model_id
+    )
     return safety_feature_extractor, safety_checker
 
 
@@ -32,13 +140,28 @@ def monkeypatch_safety_cosine_distance():
     safety_checker_mod.cosine_distance = cosine_distance_float32
 
 
-def is_nsfw(img):
+_CONCEPT_DESCRIPTIONS = []
+_SPECIAL_CARE_DESCRIPTIONS = []
+
+
+def create_safety_score(img, safety_mode=SafetyMode.STRICT):
     safety_feature_extractor, safety_checker = safety_models()
     safety_checker_input = safety_feature_extractor([img], return_tensors="pt")
     clip_input = safety_checker_input.pixel_values
 
-    _, has_nsfw_concept = safety_checker(
-        images=[np.empty((2, 2))], clip_input=clip_input
-    )
+    safety_result = safety_checker(clip_input)[0]
+
+    if safety_result.is_special_care_nsfw:
+        img.paste((150, 0, 0), (0, 0, img.size[0], img.size[1]))
+        safety_result.is_filtered = True
+        logger.info(
+            f"    ⚠️🔞️  Filtering NSFW image. nsfw score: {safety_result.nsfw_score}"
+        )
+    elif safety_mode == SafetyMode.STRICT and safety_result.is_nsfw:
+        img.paste((50, 0, 0), (0, 0, img.size[0], img.size[1]))
+        safety_result.is_filtered = True
+        logger.info(
+            f"    ⚠️  Filtering NSFW image. nsfw score: {safety_result.nsfw_score}"
+        )
 
-    return has_nsfw_concept[0]
+    return safety_result

imaginairy/schema.py

@@ -191,6 +191,7 @@ class ImagineResult:
         img,
         prompt: ImaginePrompt,
         is_nsfw,
+        safety_score,
         upscaled_img=None,
         modified_original=None,
         mask_binary=None,
@@ -217,6 +218,7 @@ class ImagineResult:
         self.upscaled_img = upscaled_img
 
         self.is_nsfw = is_nsfw
+        self.safety_score = safety_score
         self.created_at = datetime.utcnow().replace(tzinfo=timezone.utc)
         self.torch_backend = get_device()
         self.hardware_name = get_hardware_description(get_device())

tests/test_enhancers.py

@@ -146,7 +146,10 @@ def test_clip_mask_parser(mask_text, expected):
 def test_describe_picture():
     img = Image.open(f"{TESTS_FOLDER}/data/girl_with_a_pearl_earring.jpg")
     caption = generate_caption(img)
-    assert caption == "a painting of a girl with a pearl ear"
+    assert (
+        caption
+        == "a painting of a girl with a pearl earring wearing a yellow dress and a pearl earring in her ear and a black background"
+    )
 
 
 def test_clip_text_comparison():

tests/test_safety.py

@@ -1,13 +1,15 @@
 from PIL import Image
 
-from imaginairy.safety import is_nsfw
+from imaginairy.safety import create_safety_score
 from tests import TESTS_FOLDER
 
 
 def test_is_nsfw():
     img = Image.open(f"{TESTS_FOLDER}/data/safety.jpg")
 
-    assert is_nsfw(img)
+    safety_score = create_safety_score(img)
+    assert safety_score.is_nsfw
 
     img = Image.open(f"{TESTS_FOLDER}/data/girl_with_a_pearl_earring.jpg")
-    assert not is_nsfw(img)
+    safety_score = create_safety_score(img)
+    assert not safety_score.is_nsfw