fix marigold

Signed-off-by: Vladimir Mandic <mandic00@live.com>

CHANGELOG.md

@@ -39,6 +39,7 @@ TBD
   - ui: fix collapsible panels  
   - process: improve send-to functionality  
   - control: safe load non-sparse controlnet  
+  - control: fix marigold preprocessor with bfloat16  
   - auth: fix password being shown in clear text during login  
 
 ## Update for 2025-11-06

modules/control/proc/marigold/__init__.py

@@ -1,3 +1,4 @@
+import torch
 from PIL import Image
 from modules.control.util import HWC3, resize_image
 from modules import devices
@@ -28,7 +29,7 @@ class MarigoldDetector:
         color_map: str = "Spectral",
         output_type=None,
     ):
-        self.model.to(device=devices.device, dtype=devices.dtype)
+        self.model.to(device=devices.device, dtype=torch.float16)
         res = self.model(
             input_image,
             denoising_steps=denoising_steps,

modules/control/proc/marigold/marigold_pipeline.py

@@ -228,7 +228,7 @@ class MarigoldPipeline(DiffusionPipeline):
         depth_pred = (depth_pred - min_d) / (max_d - min_d)
 
         # Convert to numpy
-        depth_pred = depth_pred.cpu().numpy().astype(np.float32)
+        depth_pred = depth_pred.to(torch.float32).cpu().numpy()
 
         # Resize back to original resolution
         if match_input_res:

modules/control/proc/marigold/util/ensemble.py

@@ -64,8 +64,9 @@ def ensemble_depths(
             input_images = downscaler(torch.from_numpy(input_images)).numpy()
 
     # init guess
-    _min = np.min(input_images.reshape((n_img, -1)).cpu().numpy(), axis=1)
-    _max = np.max(input_images.reshape((n_img, -1)).cpu().numpy(), axis=1)
+    np_img = input_images.reshape((n_img, -1)).to(torch.float32).cpu().numpy()
+    _min = np.min(np_img, axis=1)
+    _max = np.max(np_img, axis=1)
     s_init = 1.0 / (_max - _min).reshape((-1, 1, 1))
     t_init = (-1 * s_init.flatten() * _min.flatten()).reshape((-1, 1, 1))
     x = np.concatenate([s_init, t_init]).reshape(-1).astype(np_dtype)
@@ -95,7 +96,7 @@ def ensemble_depths(
         far_err = torch.sqrt((1 - torch.max(pred)) ** 2)
 
         err = sqrt_dist + (near_err + far_err) * regularizer_strength
-        err = err.detach().cpu().numpy().astype(np_dtype)
+        err = err.to(torch.float32).detach().cpu().numpy().astype(np_dtype)
         return err
 
     res = minimize(