Merge branch 'dev'

CHANGELOG.md

@@ -1,7 +1,15 @@
 # Changelog
 
+## 2024-03-21
+
+- v24.3.2
+- UI를 거치지 않은 입력에 대해, image_mask를 입력했을 때 opencv 에러가 발생하는 것 수정
+- img2img inpaint에서 skip img2img 옵션을 활성화할 경우, adetailer를 비활성화함
+  - 마스크 크기에 대해 해결하기 힘든 문제가 있음
+
 ## 2024-03-16
 
+- v24.3.1
 - YOLO World v2, YOLO9 지원가능한 버전으로 ultralytics 업데이트
 - inpaint full res인 경우 인페인트 모드에서 동작하게 변경
 - inpaint full res가 아닌 경우, 사용자가 입력한 마스크와 교차점이 있는 마스크만 선택하여 사용함

adetailer/__version__.py

@@ -1 +1 @@
-__version__ = "24.3.1"
+__version__ = "24.3.2"

adetailer/mask.py

@@ -86,7 +86,7 @@ def offset(img: Image.Image, x: int = 0, y: int = 0) -> Image.Image:
 
 def is_all_black(img: Image.Image | np.ndarray) -> bool:
     if isinstance(img, Image.Image):
-        img = np.array(img)
+        img = np.array(ensure_pil_image(img, "L"))
     return cv2.countNonZero(img) == 0
 
 

pyproject.toml

@@ -7,7 +7,7 @@ readme = "README.md"
 license = { text = "AGPL-3.0" }
 dependencies = [
     "ultralytics>=8.1",
-    "mediapipe>=10",
+    "mediapipe>=0.10",
     "pydantic<3",
     "rich>=13",
     "huggingface_hub",
@@ -18,6 +18,10 @@ keywords = [
     "adetailer",
     "ultralytics",
 ]
+classifiers = [
+    "License :: OSI Approved :: GNU Affero General Public License v3",
+    "Topic :: Scientific/Engineering :: Image Recognition",
+]
 dynamic = ["version"]
 
 [project.urls]

scripts/!adetailer.py

@@ -50,6 +50,7 @@
 from modules.processing import (
     Processed,
     StableDiffusionProcessingImg2Img,
+    create_binary_mask,
     create_infotext,
     process_images,
 )
@@ -200,7 +201,7 @@ def is_ad_enabled(self, *args_) -> bool:
         not_none = any(arg.get("ad_model", "None") != "None" for arg in arg_list)
         return ad_enabled and not_none
 
-    def check_skip_img2img(self, p, *args_) -> None:
+    def set_skip_img2img(self, p, *args_) -> None:
         if (
             hasattr(p, "_ad_skip_img2img")
             or not hasattr(p, "init_images")
@@ -210,20 +211,29 @@ def check_skip_img2img(self, p, *args_) -> None:
 
         if len(args_) >= 2 and isinstance(args_[1], bool):
             p._ad_skip_img2img = args_[1]
-            if args_[1]:
-                p._ad_orig = SkipImg2ImgOrig(
-                    steps=p.steps,
-                    sampler_name=p.sampler_name,
-                    width=p.width,
-                    height=p.height,
-                )
-                p.steps = 1
-                p.sampler_name = "Euler"
-                p.width = 128
-                p.height = 128
         else:
             p._ad_skip_img2img = False
 
+        if not p._ad_skip_img2img:
+            return
+
+        if self.is_img2img_inpaint(p):
+            p._ad_disabled = True
+            msg = "[-] ADetailer: img2img inpainting with skip img2img is not supported. (because it's buggy)"
+            print(msg)
+            return
+
+        p._ad_orig = SkipImg2ImgOrig(
+            steps=p.steps,
+            sampler_name=p.sampler_name,
+            width=p.width,
+            height=p.height,
+        )
+        p.steps = 1
+        p.sampler_name = "Euler"
+        p.width = 128
+        p.height = 128
+
     @staticmethod
     def get_i(p) -> int:
         it = p.iteration
@@ -577,10 +587,10 @@ def pred_preprocessing(self, p, pred: PredictOutput, args: ADetailerArgs):
             y_offset=args.ad_y_offset,
             merge_invert=args.ad_mask_merge_invert,
         )
+
         if self.is_img2img_inpaint(p) and not self.is_inpaint_only_masked(p):
-            invert = p.inpainting_mask_invert
-            image_mask = ensure_pil_image(p.image_mask, mode="L")
-            masks = self.inpaint_mask_filter(image_mask, masks, invert)
+            image_mask = self.get_image_mask(p)
+            masks = self.inpaint_mask_filter(image_mask, masks)
         return masks
 
     @staticmethod
@@ -641,40 +651,56 @@ def is_inpaint_only_masked(p) -> bool:
 
     @staticmethod
     def inpaint_mask_filter(
-        img2img_mask: Image.Image, ad_mask: list[Image.Image], invert: int = 0
+        img2img_mask: Image.Image, ad_mask: list[Image.Image]
     ) -> list[Image.Image]:
-        if invert:
-            img2img_mask = ImageChops.invert(img2img_mask)
         return [mask for mask in ad_mask if has_intersection(img2img_mask, mask)]
 
+    @staticmethod
+    def get_image_mask(p) -> Image.Image:
+        mask = p.image_mask
+        if p.inpainting_mask_invert:
+            mask = ImageChops.invert(mask)
+        mask = create_binary_mask(mask)
+
+        if getattr(p, "_ad_skip_img2img", False):
+            width, height = p.init_images[0].size
+        else:
+            width, height = p.width, p.height
+        return images.resize_image(p.resize_mode, mask, width, height)
+
     @rich_traceback
     def process(self, p, *args_):
         if getattr(p, "_ad_disabled", False):
             return
 
-        if self.is_img2img_inpaint(p) and is_all_black(p.image_mask):
+        if self.is_img2img_inpaint(p) and is_all_black(self.get_image_mask(p)):
             p._ad_disabled = True
             msg = (
                 "[-] ADetailer: img2img inpainting with no mask -- adetailer disabled."
             )
             print(msg)
             return
 
-        if self.is_ad_enabled(*args_):
-            arg_list = self.get_args(p, *args_)
-            self.check_skip_img2img(p, *args_)
+        if not self.is_ad_enabled(*args_):
+            p._ad_disabled = True
+            return
 
-            if hasattr(p, "_ad_xyz_prompt_sr"):
-                replaced_positive_prompt, replaced_negative_prompt = self.get_prompt(
-                    p, arg_list[0]
-                )
-                arg_list[0].ad_prompt = replaced_positive_prompt[0]
-                arg_list[0].ad_negative_prompt = replaced_negative_prompt[0]
+        self.set_skip_img2img(p, *args_)
+        if getattr(p, "_ad_disabled", False):
+            # case when img2img inpainting with skip img2img
+            return
 
-            extra_params = self.extra_params(arg_list)
-            p.extra_generation_params.update(extra_params)
-        else:
-            p._ad_disabled = True
+        arg_list = self.get_args(p, *args_)
+
+        if hasattr(p, "_ad_xyz_prompt_sr"):
+            replaced_positive_prompt, replaced_negative_prompt = self.get_prompt(
+                p, arg_list[0]
+            )
+            arg_list[0].ad_prompt = replaced_positive_prompt[0]
+            arg_list[0].ad_negative_prompt = replaced_negative_prompt[0]
+
+        extra_params = self.extra_params(arg_list)
+        p.extra_generation_params.update(extra_params)
 
     def _postprocess_image_inner(
         self, p, pp, args: ADetailerArgs, *, n: int = 0

tests/test_mask.py

@@ -1,4 +1,6 @@
+import cv2
 import numpy as np
+import pytest
 from PIL import Image, ImageDraw
 
 from adetailer.mask import dilate_erode, has_intersection, is_all_black, offset
@@ -78,6 +80,24 @@ def test_is_all_black_2():
     assert not is_all_black(img)
 
 
+def test_is_all_black_rgb_image_pil():
+    img = Image.new("RGB", (10, 10), color="red")
+    assert not is_all_black(img)
+
+    img = Image.new("RGBA", (10, 10), color="red")
+    assert not is_all_black(img)
+
+
+def test_is_all_black_rgb_image_numpy():
+    img = np.full((10, 10, 4), 127, dtype=np.uint8)
+    with pytest.raises(cv2.error):
+        is_all_black(img)
+
+    img = np.full((4, 10, 10), 0.5, dtype=np.float32)
+    with pytest.raises(cv2.error):
+        is_all_black(img)
+
+
 def test_has_intersection_1():
     arr1 = np.array(
         [

tests/test_ultralytics.py

@@ -11,8 +11,10 @@
         "face_yolov8n.pt",
         "face_yolov8n_v2.pt",
         "face_yolov8s.pt",
+        "face_yolov9c.pt",
         "hand_yolov8n.pt",
         "hand_yolov8s.pt",
+        "hand_yolov9c.pt",
         "person_yolov8n-seg.pt",
         "person_yolov8s-seg.pt",
         "person_yolov8m-seg.pt",