diff --git a/libs/spandrel/spandrel/architectures/PLKSR/arch/PLKSR.py b/libs/spandrel/spandrel/architectures/PLKSR/arch/PLKSR.py
index b030271e..08094412 100644
--- a/libs/spandrel/spandrel/architectures/PLKSR/arch/PLKSR.py
+++ b/libs/spandrel/spandrel/architectures/PLKSR/arch/PLKSR.py
@@ -64,11 +64,19 @@ def __init__(self, dim, kernel_size, with_idt):
         self.idx = dim
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
-        if self.with_idt:
-            x[:, : self.idx] = x[:, : self.idx] + self.conv(x[:, : self.idx])
+        if self.training:
+            x1, x2 = torch.split(x, [self.idx, x.size(1) - self.idx], dim=1)
+            if self.with_idt:
+                x1 = self.conv(x1) + x1
+            else:
+                x1 = self.conv(x1)
+            return torch.cat([x1, x2], dim=1)
         else:
-            x[:, : self.idx] = self.conv(x[:, : self.idx])
-        return x
+            if self.with_idt:
+                x[:, : self.idx] = x[:, : self.idx] + self.conv(x[:, : self.idx])
+            else:
+                x[:, : self.idx] = self.conv(x[:, : self.idx])
+            return x
 
 
 class RectSparsePLKConv2d(nn.Module):
diff --git a/libs/spandrel/spandrel/architectures/PLKSR/arch/RealPLKSR.py b/libs/spandrel/spandrel/architectures/PLKSR/arch/RealPLKSR.py
index e8b078c1..0501d535 100644
--- a/libs/spandrel/spandrel/architectures/PLKSR/arch/RealPLKSR.py
+++ b/libs/spandrel/spandrel/architectures/PLKSR/arch/RealPLKSR.py
@@ -29,6 +29,10 @@ def __init__(self, dim: int, kernel_size: int):
         self.idx = dim
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.training:
+            x1, x2 = torch.split(x, [self.idx, x.size(1) - self.idx], dim=1)
+            x1 = self.conv(x1)
+            return torch.cat([x1, x2], dim=1)
         x[:, : self.idx] = self.conv(x[:, : self.idx])
         return x
 
@@ -109,7 +113,9 @@ def __init__(
         dropout: float = 0,
     ):
         super().__init__()
-        dropout = 0
+
+        if not self.training:
+            dropout = 0
 
         self.feats = nn.Sequential(
             *[nn.Conv2d(3, dim, 3, 1, 1)]
diff --git a/libs/spandrel/spandrel/architectures/SPAN/arch/span.py b/libs/spandrel/spandrel/architectures/SPAN/arch/span.py
index e24318f1..503201b1 100644
--- a/libs/spandrel/spandrel/architectures/SPAN/arch/span.py
+++ b/libs/spandrel/spandrel/architectures/SPAN/arch/span.py
@@ -145,9 +145,11 @@ def __init__(
             stride=s,
             bias=bias,
         )
-        self.eval_conv.weight.requires_grad = False
-        self.eval_conv.bias.requires_grad = False  # type: ignore
-        self.update_params()
+
+        if not self.training:
+            self.eval_conv.weight.requires_grad = False
+            self.eval_conv.bias.requires_grad = False  # type: ignore
+            self.update_params()
 
     def update_params(self):
         w1 = self.conv[0].weight.data.clone().detach()