fix: normalize transparent image inputs

Author: hgmzhn

doc/CHANGELOG_v2.2.10.md

@@ -55,6 +55,7 @@
 - 修复 API 管理面板图片类接口测试图过小的问题；测试用 PNG 从 2 像素调整为 50 像素，避免部分图片 API 拒绝过小输入。
 - 修复开启气泡英文排版时，渲染阶段尝试写入只读的 `TextBlock.horizontal` 属性导致任务中断的问题；现在会通过文本方向字段强制横排。
 - 修复 AI 渲染默认提示词只要求去掉编号框、未明确要求去掉数字标签的问题；启动时会自动迁移旧版默认渲染提示词，避免保留编号数字。
+- 修复带透明通道或非标准 4 通道模式图片在 Real-CUGAN、MangaJaNai、ESRGAN、AI 上色/渲染、OCR 等路径中可能被作为 4 通道张量输入模型的问题；现在统一在图像边界将 `RGBA`、`RGBa`、`LA`、带透明 `P`、`RGBX`、`CMYK` 等模式规范化为白底 RGB。
 
 ## 📝 说明
 

manga_translator/colorization/manga_colorization_v2.py

@@ -6,6 +6,7 @@
 from PIL import Image
 from torchvision.transforms import ToTensor
 
+from ..utils.image_modes import normalize_rgb_image
 from .common import OfflineColorizer
 from .manga_colorization_v2_utils.denoising.denoiser import FFDNetDenoiser
 from .manga_colorization_v2_utils.networks.models import Colorizer
@@ -49,7 +50,7 @@ async def _unload(self):
 
     async def _infer(self, image: Image.Image, colorization_size: int, denoise_sigma=25, **kwargs) -> Image.Image:
         # Size has to be multiple of 32
-        img = np.array(image.convert('RGBA'))
+        img = np.array(normalize_rgb_image(image))
         original_size = image.size  # 保存原始尺寸 (width, height)
         max_size = min(*img.shape[:2])
         max_size -= max_size % 32

manga_translator/colorization/model_api_colorizer.py

@@ -169,7 +169,7 @@ def _build_colorizer_request(self, image: Image.Image, kwargs) -> tuple[str, lis
                 continue
             try:
                 with Image.open(ref.resolved_path) as reference_image:
-                    reference_rgb = reference_image.convert("RGB")
+                    reference_rgb = normalize_ai_image(reference_image)
                 reference_images.append(
                     {
                         "kind": "prompt_reference",
@@ -269,10 +269,10 @@ async def _fetch_image_from_url(self, url: str) -> Image.Image:
         response = await self.client.session.get(url, timeout=600.0)
         if response.status_code != 200:
             raise RuntimeError(f"Failed to download generated image: HTTP {response.status_code}")
-        return Image.open(io.BytesIO(response.content)).convert("RGB")
+        return normalize_ai_image(Image.open(io.BytesIO(response.content)))
 
     def _load_image_from_bytes(self, payload: bytes) -> Image.Image:
-        return Image.open(io.BytesIO(payload)).convert("RGB")
+        return normalize_ai_image(Image.open(io.BytesIO(payload)))
 
     async def _reset_client_for_candidate(self, endpoint, error: Exception):
         del endpoint, error

manga_translator/detection/panel_finder.py

@@ -5,6 +5,8 @@
 import numpy as np
 from PIL import Image
 
+from ..utils.image_modes import normalize_rgb_image
+
 KERNEL_SIZE = 7
 BORDER_SIZE = 10
 
@@ -19,7 +21,7 @@ def panel_process_image(img: Image.Image):
         The preprocessed image.
     """
 
-    img_gray = cv.cvtColor(np.array(img), cv.COLOR_BGR2GRAY)
+    img_gray = cv.cvtColor(np.array(normalize_rgb_image(img)), cv.COLOR_RGB2GRAY)
     img_gray = cv.GaussianBlur(img_gray, (KERNEL_SIZE, KERNEL_SIZE), 0)
     img_gray = cv.threshold(img_gray, 200, 255, cv.THRESH_BINARY)[1]
 

manga_translator/inpainting/booru_tagger.py

@@ -6,6 +6,7 @@
 import pandas as pd
 from PIL import Image
 
+from ..utils.image_modes import normalize_rgb_image
 from ..utils.onnx_runtime import (
     create_inference_session,
     create_session_options,
@@ -58,11 +59,7 @@ def __init__(self, filename) -> None:
         _, self.height, _, _ = self.model.get_inputs()[0].shape
 
     def label(self, image: Image) -> Dict[str, float] :
-        # alpha to white
-        image = image.convert('RGBA')
-        new_image = Image.new('RGBA', image.size, 'WHITE')
-        new_image.paste(image, mask=image)
-        image = new_image.convert('RGB')
+        image = normalize_rgb_image(image)
         image = np.asarray(image)
 
         # PIL RGB to OpenCV BGR

manga_translator/manga_translator.py

@@ -81,6 +81,7 @@
     get_original_txt_path,
     get_work_image_path,
 )
+from .utils.ai_image_preprocess import normalize_ai_image
 from .utils.translation_text import remove_trailing_period_if_needed
 
 # Will be overwritten by __main__.py if module is being run directly (with python -m)
@@ -1542,7 +1543,7 @@ def _append_colorizer_history_image(self, config: Config, image) -> None:
         if not isinstance(image, Image.Image):
             image = Image.fromarray(np.asarray(image).astype(np.uint8))
 
-        history_image = image.convert("RGB").copy()
+        history_image = normalize_ai_image(image).copy()
         self._colorizer_history_images.append(history_image)
 
         history_pages = self._get_ai_colorizer_history_pages(config)

manga_translator/ocr/model_api_ocr.py

@@ -20,6 +20,7 @@
 from ..api_key_rotation import run_with_api_candidates
 from ..utils import Quadrilateral
 from ..utils.generic import AvgMeter
+from ..utils.image_modes import normalize_rgb_image
 from .common import OfflineOCR
 from .prompt_loader import (
     DEFAULT_AI_OCR_PROMPT,
@@ -175,8 +176,7 @@ def _normalize_ocr_text(self, text: str) -> str:
 
     def _encode_region_png_base64(self, region: np.ndarray) -> str:
         image = Image.fromarray(region.astype(np.uint8))
-        if image.mode != "RGB":
-            image = image.convert("RGB")
+        image = normalize_rgb_image(image)
         buffer = io.BytesIO()
         image.save(buffer, format="PNG")
         return base64.b64encode(buffer.getvalue()).decode("ascii")

manga_translator/ocr/model_paddleocr_vl.py

@@ -19,6 +19,7 @@
 from ..config import OcrConfig
 from ..utils import Quadrilateral
 from ..utils.generic import AvgMeter
+from ..utils.image_modes import normalize_rgb_image
 from .common import OfflineOCR
 
 
@@ -398,9 +399,7 @@ def _recognize_single(self, img: np.ndarray, prompt_text: str) -> str:
         else:
             pil_img = img
 
-        # 确保是 RGB 模式
-        if pil_img.mode != 'RGB':
-            pil_img = pil_img.convert('RGB')
+        pil_img = normalize_rgb_image(pil_img)
 
         last_output = ''
         for attempt, (generation_config, retry_suffix) in enumerate(zip(

manga_translator/rendering/model_api_renderer.py

@@ -219,7 +219,7 @@ async def _fetch_image_from_url(self, url: str) -> Image.Image:
         response = await self.client.session.get(url, timeout=600.0)
         if response.status_code != 200:
             raise RuntimeError(f"Failed to download rendered image: HTTP {response.status_code}")
-        return Image.open(io.BytesIO(response.content)).convert("RGB")
+        return normalize_ai_image(Image.open(io.BytesIO(response.content)))
 
     async def _reset_client_for_candidate(self, endpoint, error: Exception):
         del endpoint, error
@@ -253,7 +253,7 @@ def _extract_gemini_image(self, response) -> Optional[Image.Image]:
                 data = inline_data.get("data")
                 if not data:
                     continue
-                return Image.open(io.BytesIO(base64.b64decode(data))).convert("RGB")
+                return normalize_ai_image(Image.open(io.BytesIO(base64.b64decode(data))))
         return None
 
     async def render(self, img: np.ndarray, text_regions: List[TextBlock], config) -> np.ndarray:

manga_translator/rendering/text_render_pillow_eng.py

@@ -6,6 +6,7 @@
 from PIL import Image, ImageDraw, ImageFont
 
 from ..utils import TextBlock
+from ..utils.image_modes import normalize_rgb_image
 from .ballon_extractor import extract_ballon_region
 from .text_render_eng import _write_region_br_from_lines, seg_eng
 
@@ -254,7 +255,7 @@ def calculate_font_values(font, words, delimiter=' '):
             bboxes[i][0][j] += int(offset[j])
 
     # Apply strokes and paste text
-    img_pil = img_pil.convert("RGB")
+    img_pil = normalize_rgb_image(img_pil)
     img_array = np.array(img_pil)
 
     for rotated_layer, bbox, sw in zip(rotated_text_layers, bboxes, sws):