fix(backend): nsfw checker on non-rgb images

invokeai/backend/image_util/safety_checker.py

@@ -56,16 +56,25 @@ def has_nsfw_concept(cls, image: Image.Image) -> bool:
  features = cls.feature_extractor([image], return_tensors="pt")
  features.to(device)
  cls.safety_checker.to(device)
- x_image = np.array(image).astype(np.float32) / 255.0
+
+ # Only RGB(A) images are supported, so to prevent an error when a NSFW concept is detect in an image with
+ # a different image mode, we _must_ convert it to RGB and then to a normalized, batched np array.
+ rgb_image = image.convert("RGB")
+ # Convert to normalized (0-1) np array
+ x_image = np.array(rgb_image).astype(np.float32) / 255.0
+ # Add batch dimension and transpose to NCHW
  x_image = x_image[None].transpose(0, 3, 1, 2)
+ # A warning is logged if a NSFW concept is detected - silence those, so we can handle it ourselves.
  with SilenceWarnings():
+ # `clip_input` (features) is used to check for NSFW concepts. `images` is required, but it isn't actually
+ # checked for NSFW concepts. If a NSFW concept is detected, the the image is replaced with a black image.
  checked_image, has_nsfw_concept = cls.safety_checker(images=x_image, clip_input=features.pixel_values)
  return has_nsfw_concept[0]
 
  @classmethod
  def blur_if_nsfw(cls, image: Image.Image) -> Image.Image:
  if cls.has_nsfw_concept(image):
- logger.info("A potentially NSFW image has been detected. Image will be blurred.")
+ logger.warning("A potentially NSFW image has been detected. Image will be blurred.")
  blurry_image = image.filter(filter=ImageFilter.GaussianBlur(radius=32))
  caution = cls._get_caution_img()
  # Center the caution image on the blurred image