From 9866d770a6a75eb0a9ca0d1f3a31154a07a9024b Mon Sep 17 00:00:00 2001
From: qinxuye <qinxuye@gmail.com>
Date: Tue, 1 Oct 2024 15:55:17 +0800
Subject: [PATCH] Fix adetailer

---
 xinference/model/image/adetailer.py             | 11 ++++++++++-
 xinference/model/image/sdapi.py                 |  4 +---
 xinference/model/image/stable_diffusion/core.py |  7 +++----
 3 files changed, 14 insertions(+), 8 deletions(-)

diff --git a/xinference/model/image/adetailer.py b/xinference/model/image/adetailer.py
index 359923b24d..b8393662af 100644
--- a/xinference/model/image/adetailer.py
+++ b/xinference/model/image/adetailer.py
@@ -277,6 +277,10 @@ def pred_preprocessing(pred: PredictOutput, args: ADetailerArgs):
             new_kwargs["image"] = ensure_pil_image(init_image, "RGB")
             denoising_strength = new_kwargs.get("strength")
             new_kwargs["strength"] = get_dynamic_denoise_strength(denoising_strength, pred.bboxes[j], image.size)  # type: ignore
+            if adetailer.ad_inpaint_only_masked:
+                new_kwargs[
+                    "padding_mask_crop"
+                ] = adetailer.ad_inpaint_only_masked_padding
 
             # Don't override user-defined dimensions.
             if not adetailer.ad_use_inpaint_width_height:
@@ -284,7 +288,12 @@ def pred_preprocessing(pred: PredictOutput, args: ADetailerArgs):
                     width, height, pred.bboxes[j]
                 )
 
-            init_image = model.image_to_image(**new_kwargs)[0]  # type: ignore
+            if "width" in new_kwargs and "size" not in new_kwargs:
+                new_kwargs[
+                    "size"
+                ] = f"{new_kwargs.pop('width')} * {new_kwargs.pop('height')}"
+
+            init_image = model.inpainting(**new_kwargs)[0]  # type: ignore
 
         image = init_image
 
diff --git a/xinference/model/image/sdapi.py b/xinference/model/image/sdapi.py
index 1b28701678..83295eab8e 100644
--- a/xinference/model/image/sdapi.py
+++ b/xinference/model/image/sdapi.py
@@ -192,9 +192,7 @@ def txt2img(self, **kwargs):
                 "created": int(time.time()),
             }
             for i, result_image in enumerate(result_images):
-                result["data"][i]["b64_json"] = encode_pil_to_base64(
-                    result_image
-                ).decode()
+                result["data"][i]["b64_json"] = encode_pil_to_base64(result_image)
 
         # convert to SD API result
         return {
diff --git a/xinference/model/image/stable_diffusion/core.py b/xinference/model/image/stable_diffusion/core.py
index d55f60b062..abc82cd354 100644
--- a/xinference/model/image/stable_diffusion/core.py
+++ b/xinference/model/image/stable_diffusion/core.py
@@ -530,7 +530,8 @@ def inpainting(
             logger.debug("Process mask image with mask_blur: %s", mask_blur)
             mask_image = model.mask_processor.blur(mask_image, blur_factor=mask_blur)  # type: ignore
 
-        width, height = map(int, re.split(r"[^\d]+", size))
+        if "width" not in kwargs:
+            kwargs["width"], kwargs["height"] = map(int, re.split(r"[^\d]+", size))
 
         if padding_image_to_multiple := kwargs.pop("padding_image_to_multiple", None):
             # Model like SD3 inpainting requires image's height and width is times of 16
@@ -543,14 +544,12 @@ def inpainting(
                 mask_image, multiple=int(padding_image_to_multiple)
             )
             # calculate actual image size after padding
-            width, height = image.size
+            kwargs["width"], kwargs["height"] = image.size
 
         return self._call_model(
             image=image,
             mask_image=mask_image,
             prompt=prompt,
-            height=height,
-            width=width,
             num_images_per_prompt=n,
             response_format=response_format,
             model=model,