diff --git a/extensions-builtin/Lora/network_oft.py b/extensions-builtin/Lora/network_oft.py
index 2be67fe53..93402bb28 100644
--- a/extensions-builtin/Lora/network_oft.py
+++ b/extensions-builtin/Lora/network_oft.py
@@ -24,12 +24,14 @@ class NetworkModuleOFT(network.NetworkModule):
         # kohya-ss
         if "oft_blocks" in weights.w.keys():
             self.is_kohya = True
-            self.oft_blocks = weights.w["oft_blocks"]
+            self.oft_blocks = weights.w["oft_blocks"] # (num_blocks, block_size, block_size)
             self.alpha = weights.w["alpha"]
-            self.dim = self.oft_blocks.shape[0]
+            self.dim = self.oft_blocks.shape[0] # lora dim
+            #self.oft_blocks = rearrange(self.oft_blocks, 'k m ... -> (k m) ...')
         elif "oft_diag" in weights.w.keys():
             self.is_kohya = False
-            self.oft_blocks = weights.w["oft_diag"]
+            self.oft_blocks = weights.w["oft_diag"] # (num_blocks, block_size, block_size)
+
             # alpha is rank if alpha is 0 or None
             if self.alpha is None:
                 pass
@@ -51,12 +53,11 @@ class NetworkModuleOFT(network.NetworkModule):
             raise ValueError("sd_module must be Linear or Conv")
 
         if self.is_kohya:
-            self.num_blocks = self.dim
-            self.block_size = self.out_dim // self.num_blocks
             self.constraint = self.alpha * self.out_dim
+            self.num_blocks, self.block_size = factorization(self.out_dim, self.dim)
         else:
-            self.block_size, self.num_blocks = factorization(self.out_dim, self.dim)
             self.constraint = None
+            self.block_size, self.num_blocks = factorization(self.out_dim, self.dim)
 
     def merge_weight(self, R_weight, org_weight):
         R_weight = R_weight.to(org_weight.device, dtype=org_weight.dtype)
@@ -77,7 +78,8 @@ class NetworkModuleOFT(network.NetworkModule):
         else:
             new_norm_Q = norm_Q
         block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8))
-        m_I = torch.eye(self.block_size, device=oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1)
+        m_I = torch.eye(self.num_blocks, device=oft_blocks.device).unsqueeze(0).repeat(self.block_size, 1, 1)
+        #m_I = torch.eye(self.block_size, device=oft_blocks.device).unsqueeze(0).repeat(self.num_blocks, 1, 1)
         block_R = torch.matmul(m_I + block_Q, (m_I - block_Q).inverse())
 
         block_R_weighted = multiplier * block_R + (1 - multiplier) * m_I
@@ -97,25 +99,33 @@ class NetworkModuleOFT(network.NetworkModule):
         is_other_linear = type(self.sd_module) in [torch.nn.MultiheadAttention]
 
         if not is_other_linear:
-            if is_other_linear and orig_weight.shape[0] != orig_weight.shape[1]:
-                orig_weight=orig_weight.permute(1, 0)
+            #if is_other_linear and orig_weight.shape[0] != orig_weight.shape[1]:
+            #    orig_weight=orig_weight.permute(1, 0)
+
+            oft_blocks = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
+
+            # without this line the results are significantly worse / less accurate
+            oft_blocks = oft_blocks - oft_blocks.transpose(1, 2)
+
+            R = oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
+            R = R * multiplier + torch.eye(self.block_size, device=orig_weight.device)
 
-            R = self.oft_blocks.to(orig_weight.device, dtype=orig_weight.dtype)
             merged_weight = rearrange(orig_weight, '(k n) ... -> k n ...', k=self.num_blocks, n=self.block_size)
             merged_weight = torch.einsum(
                 'k n m, k n ... -> k m ...',
-                R * multiplier + torch.eye(self.block_size, device=orig_weight.device),
+                R,
                 merged_weight
             )
             merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...')
 
-            if is_other_linear and orig_weight.shape[0] != orig_weight.shape[1]:
-                orig_weight=orig_weight.permute(1, 0)
+            #if is_other_linear and orig_weight.shape[0] != orig_weight.shape[1]:
+            #    orig_weight=orig_weight.permute(1, 0)
 
             updown = merged_weight.to(orig_weight.device, dtype=orig_weight.dtype) - orig_weight
             output_shape = orig_weight.shape
         else:
             # FIXME: skip MultiheadAttention for now
+            #up = self.lin_module.weight.to(orig_weight.device, dtype=orig_weight.dtype)
             updown = torch.zeros([orig_weight.shape[1], orig_weight.shape[1]], device=orig_weight.device, dtype=orig_weight.dtype)
             output_shape = (orig_weight.shape[1], orig_weight.shape[1])
 
@@ -123,10 +133,10 @@ class NetworkModuleOFT(network.NetworkModule):
 
     def calc_updown(self, orig_weight):
         multiplier = self.multiplier() * self.calc_scale()
-        if self.is_kohya:
-            return self.calc_updown_kohya(orig_weight, multiplier)
-        else:
-            return self.calc_updown_kb(orig_weight, multiplier)
+        #if self.is_kohya:
+        #    return self.calc_updown_kohya(orig_weight, multiplier)
+        #else:
+        return self.calc_updown_kb(orig_weight, multiplier)
 
     # override to remove the multiplier/scale factor; it's already multiplied in get_weight
     def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None):