make mask blur work

lama_cleaner/app/src/adapters/inpainting.ts

@@ -39,11 +39,11 @@ export default async function inpaint(
   fd.append('croperHeight', croperRect.height.toString())
   fd.append('croperWidth', croperRect.width.toString())
   fd.append('useCroper', settings.showCroper ? 'true' : 'false')
+  fd.append('sdMaskBlur', settings.sdMaskBlur.toString())
   fd.append('sdStrength', settings.sdStrength.toString())
   fd.append('sdSteps', settings.sdSteps.toString())
   fd.append('sdGuidanceScale', settings.sdGuidanceScale.toString())
   fd.append('sdSampler', settings.sdSampler.toString())
-  // fd.append('sdSeed', settings.sdSeedFixed ? settings.sdSeed.toString() : '-1')
   fd.append('sdSeed', seed ? seed.toString() : '-1')
 
   if (sizeLimit === undefined) {

lama_cleaner/app/src/store/Atoms.tsx

@@ -231,7 +231,7 @@ export const settingStateDefault: Settings = {
   zitsWireframe: true,
 
   // SD
-  sdMaskBlur: 0,
+  sdMaskBlur: 5,
   sdMode: SDMode.inpainting,
   sdStrength: 0.75,
   sdSteps: 50,

lama_cleaner/model/sd.py

@@ -48,7 +48,7 @@ class SD(InpaintModel):
 
         self.model = StableDiffusionInpaintPipeline.from_pretrained(
             self.model_id_or_path,
-            revision="fp16" if torch.cuda.is_available() else 'main',
+            revision="fp16" if torch.cuda.is_available() else "main",
             torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
             use_auth_token=kwargs["hf_access_token"],
         )
@@ -91,7 +91,7 @@ class SD(InpaintModel):
                 "beta_start": 0.00085,
                 "beta_end": 0.012,
                 "num_train_timesteps": 1000,
-                "skip_prk_steps": True
+                "skip_prk_steps": True,
             }
             scheduler = PNDMScheduler(**PNDM_kwargs)
         else:
@@ -105,6 +105,10 @@ class SD(InpaintModel):
         torch.manual_seed(seed)
         torch.cuda.manual_seed_all(seed)
 
+        if config.sd_mask_blur != 0:
+            k = 2 * config.sd_mask_blur + 1
+            mask = cv2.GaussianBlur(mask, (k, k), 0)[:, :, np.newaxis]
+
         output = self.model(
             prompt=config.prompt,
             init_image=PIL.Image.fromarray(image),

lama_cleaner/schema.py

@@ -4,19 +4,19 @@ from pydantic import BaseModel
 
 
 class HDStrategy(str, Enum):
-    ORIGINAL = 'Original'
-    RESIZE = 'Resize'
-    CROP = 'Crop'
+    ORIGINAL = "Original"
+    RESIZE = "Resize"
+    CROP = "Crop"
 
 
 class LDMSampler(str, Enum):
-    ddim = 'ddim'
-    plms = 'plms'
+    ddim = "ddim"
+    plms = "plms"
 
 
 class SDSampler(str, Enum):
-    ddim = 'ddim'
-    pndm = 'pndm'
+    ddim = "ddim"
+    pndm = "pndm"
 
 
 class Config(BaseModel):
@@ -28,7 +28,7 @@ class Config(BaseModel):
     hd_strategy_crop_trigger_size: int
     hd_strategy_resize_limit: int
 
-    prompt: str = ''
+    prompt: str = ""
     # 始终是在原图尺度上的值
     use_croper: bool = False
     croper_x: int = None
@@ -37,6 +37,7 @@ class Config(BaseModel):
     croper_width: int = None
 
     # sd
+    sd_mask_blur: int = 0
     sd_strength: float = 0.75
     sd_steps: int = 50
     sd_guidance_scale: float = 7.5

lama_cleaner/server.py

@@ -118,6 +118,7 @@ def process():
         croper_y=form["croperY"],
         croper_height=form["croperHeight"],
         croper_width=form["croperWidth"],
+        sd_mask_blur=form["sdMaskBlur"],
         sd_strength=form["sdStrength"],
         sd_steps=form["sdSteps"],
         sd_guidance_scale=form["sdGuidanceScale"],