diff --git a/burn-tensor/src/tensor/backend/autodiff/ops/detach.rs b/burn-tensor/src/tensor/backend/autodiff/ops/detach.rs
deleted file mode 100644
index 0d1914945..000000000
--- a/burn-tensor/src/tensor/backend/autodiff/ops/detach.rs
+++ /dev/null
@@ -1,9 +0,0 @@
-use crate::tensor::backend::Backend;
-use crate::tensor::{backend::autodiff::ADTensor, ops::*};
-
-impl<B: Backend, P, const D: usize> TensorOpsDetach<P, D> for ADTensor<D, B> {
-    fn detach(self) -> Self {
-        let tensor = self.tensor();
-        Self::from_tensor(tensor.detach())
-    }
-}
diff --git a/burn-tensor/src/tensor/backend/autodiff/ops/mod.rs b/burn-tensor/src/tensor/backend/autodiff/ops/mod.rs
index ffce8f406..ec40330b8 100644
--- a/burn-tensor/src/tensor/backend/autodiff/ops/mod.rs
+++ b/burn-tensor/src/tensor/backend/autodiff/ops/mod.rs
@@ -3,7 +3,6 @@ mod arg;
 mod base;
 mod cat;
 mod creation;
-mod detach;
 mod erf;
 mod exp;
 mod log;
diff --git a/burn-tensor/src/tensor/backend/autodiff/ops/tensor.rs b/burn-tensor/src/tensor/backend/autodiff/ops/tensor.rs
index ae4d21dd5..d54c02501 100644
--- a/burn-tensor/src/tensor/backend/autodiff/ops/tensor.rs
+++ b/burn-tensor/src/tensor/backend/autodiff/ops/tensor.rs
@@ -702,4 +702,10 @@ impl<B: Backend> TensorOps<ADBackendDecorator<B>> for ADBackendDecorator<B> {
     ) -> <ADBackendDecorator<B> as Backend>::BoolTensorPrimitive<D> {
         B::lower_equal_scalar(lhs.tensor_ref(), rhs)
     }
+
+    fn detach<const D: usize>(
+        tensor: &<ADBackendDecorator<B> as Backend>::TensorPrimitive<D>,
+    ) -> <ADBackendDecorator<B> as Backend>::TensorPrimitive<D> {
+        ADTensor::from_tensor(B::detach(tensor.tensor_ref()))
+    }
 }
diff --git a/burn-tensor/src/tensor/backend/base.rs b/burn-tensor/src/tensor/backend/base.rs
index 5bfdafa80..c3324f14f 100644
--- a/burn-tensor/src/tensor/backend/base.rs
+++ b/burn-tensor/src/tensor/backend/base.rs
@@ -21,7 +21,6 @@ pub trait Backend:
     type FullPrecisionBackend: Backend<Elem = Self::FullPrecisionElem, Device = Self::Device>;
     type IntegerBackend: Backend<Elem = i64, Device = Self::Device>;
     type TensorPrimitive<const D: usize>: std::ops::Add<Self::TensorPrimitive<D>, Output = Self::TensorPrimitive<D>>
-        + TensorOpsDetach<Self::Elem, D>
         + Zeros<Self::TensorPrimitive<D>>
         + Ones<Self::TensorPrimitive<D>>
         + TensorOpsPrecision<Self, D>
diff --git a/burn-tensor/src/tensor/backend/ndarray/ops/detach.rs b/burn-tensor/src/tensor/backend/ndarray/ops/detach.rs
deleted file mode 100644
index 8bf07d933..000000000
--- a/burn-tensor/src/tensor/backend/ndarray/ops/detach.rs
+++ /dev/null
@@ -1,13 +0,0 @@
-use crate::{
-    tensor::{backend::ndarray::NdArrayTensor, ops::*},
-    NdArrayElement,
-};
-
-impl<E, const D: usize> TensorOpsDetach<E, D> for NdArrayTensor<E, D>
-where
-    E: NdArrayElement,
-{
-    fn detach(self) -> Self {
-        self
-    }
-}
diff --git a/burn-tensor/src/tensor/backend/ndarray/ops/mod.rs b/burn-tensor/src/tensor/backend/ndarray/ops/mod.rs
index 06ccecabb..1b1028949 100644
--- a/burn-tensor/src/tensor/backend/ndarray/ops/mod.rs
+++ b/burn-tensor/src/tensor/backend/ndarray/ops/mod.rs
@@ -2,7 +2,6 @@ mod aggregation;
 mod arg;
 mod cat;
 mod creation;
-mod detach;
 mod erf;
 mod exp;
 mod log;
diff --git a/burn-tensor/src/tensor/backend/ndarray/tensor_ops.rs b/burn-tensor/src/tensor/backend/ndarray/tensor_ops.rs
index a4b5ddd30..a3bc85b50 100644
--- a/burn-tensor/src/tensor/backend/ndarray/tensor_ops.rs
+++ b/burn-tensor/src/tensor/backend/ndarray/tensor_ops.rs
@@ -365,6 +365,9 @@ impl<E: NdArrayElement> TensorOps<NdArrayBackend<E>> for NdArrayBackend<E> {
             array,
         }
     }
+    fn detach<const D: usize>(tensor: &NdArrayTensor<E, D>) -> NdArrayTensor<E, D> {
+        tensor.clone()
+    }
 }
 
 fn to_slice_args<const D1: usize, const D2: usize>(
diff --git a/burn-tensor/src/tensor/backend/tch/ops/detach.rs b/burn-tensor/src/tensor/backend/tch/ops/detach.rs
deleted file mode 100644
index c3e02d44b..000000000
--- a/burn-tensor/src/tensor/backend/tch/ops/detach.rs
+++ /dev/null
@@ -1,13 +0,0 @@
-use crate::{
-    tensor::{backend::tch::TchTensor, ops::*},
-    TchElement,
-};
-
-impl<E, const D: usize> TensorOpsDetach<E, D> for TchTensor<E, D>
-where
-    E: TchElement,
-{
-    fn detach(self) -> Self {
-        self
-    }
-}
diff --git a/burn-tensor/src/tensor/backend/tch/ops/mod.rs b/burn-tensor/src/tensor/backend/tch/ops/mod.rs
index 06ccecabb..1b1028949 100644
--- a/burn-tensor/src/tensor/backend/tch/ops/mod.rs
+++ b/burn-tensor/src/tensor/backend/tch/ops/mod.rs
@@ -2,7 +2,6 @@ mod aggregation;
 mod arg;
 mod cat;
 mod creation;
-mod detach;
 mod erf;
 mod exp;
 mod log;
diff --git a/burn-tensor/src/tensor/backend/tch/tensor_ops.rs b/burn-tensor/src/tensor/backend/tch/tensor_ops.rs
index fa520945e..a8910fd5e 100644
--- a/burn-tensor/src/tensor/backend/tch/tensor_ops.rs
+++ b/burn-tensor/src/tensor/backend/tch/tensor_ops.rs
@@ -318,6 +318,9 @@ impl<E: TchElement> TensorOps<TchBackend<E>> for TchBackend<E> {
             kind: TchKind::<bool>::new(),
         }
     }
+    fn detach<const D: usize>(tensor: &TchTensor<E, D>) -> TchTensor<E, D> {
+        tensor.clone()
+    }
 }
 
 fn to_tensor<const D: usize, E: TchElement>(tensor: tch::Tensor) -> TchTensor<E, D> {
diff --git a/burn-tensor/src/tensor/base.rs b/burn-tensor/src/tensor/base.rs
index 276bd7ddd..b5b7116ac 100644
--- a/burn-tensor/src/tensor/base.rs
+++ b/burn-tensor/src/tensor/base.rs
@@ -519,7 +519,7 @@ where
     /// This can be used in batchers or elsewere to ensure that previous operations are not
     /// considered in the autodiff graph.
     pub fn detach(self) -> Self {
-        Self::new(self.value.detach())
+        Self::new(B::detach(&self.value))
     }
 
     /// Unsqueeze the current tensor. Create new dimensions to fit the given size.
diff --git a/burn-tensor/src/tensor/ops/base.rs b/burn-tensor/src/tensor/ops/base.rs
index a07056cde..54ac1edab 100644
--- a/burn-tensor/src/tensor/ops/base.rs
+++ b/burn-tensor/src/tensor/ops/base.rs
@@ -169,6 +169,7 @@ pub trait TensorOps<B: Backend> {
         lhs: &B::TensorPrimitive<D>,
         rhs: &B::Elem,
     ) -> B::BoolTensorPrimitive<D>;
+    fn detach<const D: usize>(tensor: &B::TensorPrimitive<D>) -> B::TensorPrimitive<D>;
 }
 
 pub trait TensorOpsAggregation<B: Backend, const D: usize> {
@@ -206,10 +207,6 @@ pub trait TensorOpsLog<E, const D: usize> {
     fn log(&self) -> Self;
 }
 
-pub trait TensorOpsDetach<E, const D: usize> {
-    fn detach(self) -> Self;
-}
-
 pub trait TensorOpsErf<E, const D: usize> {
     fn erf(&self) -> Self;
 }