diff --git a/extensions-builtin/forge_preprocessor_inpaint/scripts/preprocessor_inpaint.py b/extensions-builtin/forge_preprocessor_inpaint/scripts/preprocessor_inpaint.py
index eb8b06a7..2127da4b 100644
--- a/extensions-builtin/forge_preprocessor_inpaint/scripts/preprocessor_inpaint.py
+++ b/extensions-builtin/forge_preprocessor_inpaint/scripts/preprocessor_inpaint.py
@@ -20,8 +20,6 @@ class PreprocessorInpaintOnly(PreprocessorInpaint):
         self.name = 'inpaint_only'
         self.image = None
         self.mask = None
-        self.latent_image = None
-        self.latent_mask = None
 
     def process_before_every_sampling(self, process, cond, *args, **kwargs):
         self.image = kwargs['cond_before_inpaint_fix'][:, 0:3]
@@ -30,14 +28,25 @@ class PreprocessorInpaintOnly(PreprocessorInpaint):
         vae = process.sd_model.forge_objects.vae
         # This is a powerful VAE with integrated memory management, bf16, and tiled fallback.
 
-        self.latent_image = vae.encode(self.image.movedim(1, -1))
-        B, C, H, W = self.latent_image.shape
+        latent_image = vae.encode(self.image.movedim(1, -1))
+        latent_image = process.sd_model.forge_objects.unet.model.latent_format.process_in(latent_image)
+
+        B, C, H, W = latent_image.shape
 
         latent_mask = self.mask
         latent_mask = torch.nn.functional.interpolate(latent_mask, size=(H * 8, W * 8), mode="bilinear").round()
-        latent_mask = torch.nn.functional.max_pool2d(latent_mask, (8, 8)).round().to(self.latent_image)
-        self.latent_mask = latent_mask
+        latent_mask = torch.nn.functional.max_pool2d(latent_mask, (8, 8)).round().to(latent_image)
 
+        unet = process.sd_model.forge_objects.unet.clone()
+
+        def post_cfg(args):
+            denoised = args['denoised']
+            denoised = denoised * latent_mask.to(denoised) + latent_image.to(denoised) * (1.0 - latent_mask.to(denoised))
+            return denoised
+
+        unet.set_model_sampler_post_cfg_function(post_cfg)
+
+        process.sd_model.forge_objects.unet = unet
         return
 
     def process_after_every_sampling(self, process, params, *args, **kwargs):