diff --git a/src/KSVD.jl b/src/KSVD.jl
index 7e37318..812238a 100644
--- a/src/KSVD.jl
+++ b/src/KSVD.jl
@@ -88,6 +88,7 @@ function ksvd(Y::AbstractMatrix{T}, n_atoms::Int, max_nnz=n_atoms÷10;
               ksvd_update_method = BatchedParallelKSVD{false, T}(; shuffle_indices=true, batch_size_per_thread=1),
               sparse_coding_method = ParallelMatchingPursuit(; max_nnz, rtol=5e-2),
               minibatch_size=nothing,
+              D_init::Union{Nothing, <:AbstractMatrix{T}} = nothing,
               # termination conditions
               maxiters::Int=100, #: The maximum number of iterations to perform. Defaults to 100.
               maxtime::Union{Nothing, <:Real}=nothing,# : The maximum time for solving the nonlinear system of equations. Defaults to nothing which means no time limit. Note that setting a time limit does have a small overhead.
@@ -104,7 +105,7 @@ function ksvd(Y::AbstractMatrix{T}, n_atoms::Int, max_nnz=n_atoms÷10;
 
     # D is a dictionary matrix that contains atoms for columns.
     @timeit_debug timer "Init dict" begin
-        D = init_dictionary(T, emb_dim, n_atoms)  # size(D) == (n, K)
+        D = (isnothing(D_init) ? init_dictionary(T, emb_dim, n_atoms) : D_init)  # size(D) == (n, K)
         @assert all(≈(1.0), norm.(eachcol(D)))
     end
     X = sparse_coding(sparse_coding_method, Y, D; timer)