diff --git a/README.md b/README.md
index b2fe09af..faef39ca 100644
--- a/README.md
+++ b/README.md
@@ -42,7 +42,7 @@ Would you like to build/train a model using Keras/Python? And would you like to
 * `Add`, `Concatenate`, `Subtract`, `Multiply`, `Average`, `Maximum`, `Minimum`, `Dot`
 * `AveragePooling1D/2D/3D`, `GlobalAveragePooling1D/2D/3D`
 * `TimeDistributed`
-* `Conv1D/2D`, `SeparableConv2D`, `DepthwiseConv2D`
+* `Conv1D/2D/3D`, `SeparableConv2D`, `DepthwiseConv2D`
 * `Conv1DTranspose`, `Conv2DTranspose`
 * `Cropping1D/2D/3D`, `ZeroPadding1D/2D/3D`, `CenterCrop`
 * `BatchNormalization`, `Dense`, `Flatten`, `Normalization`
@@ -71,7 +71,7 @@ Would you like to build/train a model using Keras/Python? And would you like to
 ### Currently not supported are the following:
 
 `Lambda` ([why](FAQ.md#why-are-lambda-layers-not-supported)),
-`Conv3D`, `ConvLSTM1D`, `ConvLSTM2D`, `Discretization`,
+`ConvLSTM1D`, `ConvLSTM2D`, `Discretization`,
 `GRUCell`, `Hashing`,
 `IntegerLookup`,
 `LocallyConnected1D`, `LocallyConnected2D`,
diff --git a/include/fdeep/convolution3d.hpp b/include/fdeep/convolution3d.hpp
index 3c3e7d99..6f546908 100644
--- a/include/fdeep/convolution3d.hpp
+++ b/include/fdeep/convolution3d.hpp
@@ -8,7 +8,9 @@
 
 #include "fdeep/common.hpp"
 
+#include "fdeep/convolution.hpp"
 #include "fdeep/filter.hpp"
+#include "fdeep/shape3.hpp"
 
 #include <algorithm>
 #include <cassert>
@@ -115,5 +117,246 @@ namespace internal {
             out_size_d4_size_t, out_height_size_t, out_width_size_t };
     }
 
+    struct convolution3d_filter_matrices {
+        tensor_shape filter_shape_;
+        std::size_t filter_count_;
+        float_vec biases_;
+        bool use_bias_;
+        tensor filter_mats_;
+    };
+
+    inline tensor dilate_tensor_3d(const shape3& dilation_rate, const tensor& in)
+    {
+        if (dilation_rate == shape3(1, 1, 1)) {
+            return in;
+        }
+        assertion(in.shape().rank() == 4, "Invalid rank for 3d dilation");
+
+        const auto in_shape = in.shape();
+        const tensor_shape dilated_shape(
+            (in_shape.size_dim_4_ - 1) * dilation_rate.size_dim_4_ + 1,
+            (in_shape.height_ - 1) * dilation_rate.height_ + 1,
+            (in_shape.width_ - 1) * dilation_rate.width_ + 1,
+            in_shape.depth_);
+        tensor result(dilated_shape, static_cast<float_type>(0));
+        for (std::size_t d4 = 0; d4 < in_shape.size_dim_4_; ++d4) {
+            for (std::size_t y = 0; y < in_shape.height_; ++y) {
+                for (std::size_t x = 0; x < in_shape.width_; ++x) {
+                    for (std::size_t z = 0; z < in_shape.depth_; ++z) {
+                        result.set_ignore_rank(tensor_pos(
+                                                   d4 * dilation_rate.size_dim_4_,
+                                                   y * dilation_rate.height_,
+                                                   x * dilation_rate.width_,
+                                                   z),
+                            in.get_ignore_rank(tensor_pos(d4, y, x, z)));
+                    }
+                }
+            }
+        }
+        return result;
+    }
+
+    inline filter dilate_filter_3d(const shape3& dilation_rate, const filter& undilated)
+    {
+        return filter(dilate_tensor_3d(dilation_rate, undilated.get_tensor()),
+            undilated.get_bias());
+    }
+
+    inline filter_vec generate_filters_3d(
+        const shape3& dilation_rate,
+        const tensor_shape& filter_shape, std::size_t k,
+        const float_vec& weights, const float_vec& bias)
+    {
+        filter_vec filters(k, filter(tensor(filter_shape, 0), 0));
+
+        assertion(!filters.empty(), "at least one filter needed");
+        const std::size_t param_count = fplus::sum(fplus::transform(
+            fplus_c_mem_fn_t(filter, volume, std::size_t), filters));
+
+        assertion(static_cast<std::size_t>(weights.size()) == param_count,
+            "invalid weight size");
+        const auto filter_param_cnt = filters.front().shape().volume();
+
+        auto filter_weights = fplus::split_every(filter_param_cnt, weights);
+        assertion(filter_weights.size() == filters.size(),
+            "invalid size of filter weights");
+        assertion(bias.size() == filters.size(), "invalid bias size");
+        auto it_filter_val = std::begin(filter_weights);
+        auto it_filter_bias = std::begin(bias);
+        for (auto& filt : filters) {
+            filt.set_params(*it_filter_val, *it_filter_bias);
+            filt = dilate_filter_3d(dilation_rate, filt);
+            ++it_filter_val;
+            ++it_filter_bias;
+        }
+
+        return filters;
+    }
+
+    inline convolution3d_filter_matrices generate_im2col_filter_matrix_3d(
+        const std::vector<filter>& filters)
+    {
+        assertion(fplus::all_the_same_on(
+                      fplus_c_mem_fn_t(filter, shape, tensor_shape), filters),
+            "all filters must have the same shape");
+
+        const auto biases = fplus::transform_convert<float_vec>(
+            fplus_c_mem_fn_t(filter, get_bias, float_type),
+            filters);
+
+        const bool use_bias = fplus::sum(biases) != static_cast<float_type>(0) || !fplus::all_the_same(biases);
+
+        const auto shape = filters.front().shape();
+
+        tensor filter_mats = tensor(
+            tensor_shape(shape.size_dim_4_, shape.height_, shape.width_, shape.depth_, filters.size()),
+            static_cast<float_type>(0));
+
+        for (std::size_t d4 = 0; d4 < shape.size_dim_4_; ++d4) {
+            for (std::size_t y = 0; y < shape.height_; ++y) {
+                for (std::size_t n = 0; n < filters.size(); ++n) {
+                    for (std::size_t x = 0; x < shape.width_; ++x) {
+                        for (std::size_t z = 0; z < shape.depth_; ++z) {
+                            filter_mats.set(tensor_pos(d4, y, x, z, n),
+                                filters[n].get(tensor_pos(d4, y, x, z)));
+                        }
+                    }
+                }
+            }
+        }
+
+        return { shape, filters.size(), biases, use_bias, filter_mats };
+    }
+
+    inline tensor init_conv_output_tensor_3d(
+        std::size_t out_size_d4,
+        std::size_t out_height,
+        std::size_t out_width,
+        std::size_t out_depth,
+        std::size_t rank,
+        const convolution3d_filter_matrices& filter_mat)
+    {
+        tensor output(tensor_shape_with_changed_rank(
+                          tensor_shape(out_size_d4, out_height, out_width, out_depth),
+                          rank),
+            static_cast<float_type>(0));
+        if (filter_mat.use_bias_) {
+            const auto bias_ptr = &filter_mat.biases_.front();
+            const auto bias_ptr_end = bias_ptr + out_depth;
+            for (std::size_t d4_out = 0; d4_out < out_size_d4; ++d4_out) {
+                for (std::size_t y_out = 0; y_out < out_height; ++y_out) {
+                    for (std::size_t x_out = 0; x_out < out_width; ++x_out) {
+                        auto output_ptr = &output.get_ref_ignore_rank(tensor_pos(0, d4_out, y_out, x_out, 0));
+                        std::copy(bias_ptr, bias_ptr_end, output_ptr);
+                    }
+                }
+            }
+        }
+        return output;
+    }
+
+    inline Eigen::Map<ColMajorMatrixXf, Eigen::Unaligned, Eigen::OuterStride<>> get_im2col_mapping_3d(
+        const tensor& in,
+        std::size_t f_width,
+        std::size_t f_depth,
+        std::size_t strides_x,
+        std::size_t out_width,
+        std::size_t d4,
+        std::size_t y,
+        std::size_t d4_filt,
+        std::size_t y_filt)
+    {
+        // Same trick as in the 2D case: avoid materializing the im2col matrix
+        // by using an outer stride smaller than the row count, so adjacent
+        // columns share data along the receptive field.
+        return Eigen::Map<ColMajorMatrixXf, Eigen::Unaligned, Eigen::OuterStride<>>(
+            const_cast<float_type*>(&in.get_ref_ignore_rank(tensor_pos(0, d4 + d4_filt, y + y_filt, 0, 0))),
+            static_cast<EigenIndex>(f_width * f_depth),
+            static_cast<EigenIndex>(out_width),
+            Eigen::OuterStride<>(static_cast<EigenIndex>(f_depth * strides_x)));
+    }
+
+    inline tensor convolve_accumulative_3d(
+        std::size_t out_size_d4,
+        std::size_t out_height,
+        std::size_t out_width,
+        std::size_t strides_d4,
+        std::size_t strides_y,
+        std::size_t strides_x,
+        const convolution3d_filter_matrices& filter_mat,
+        const tensor& in)
+    {
+        const tensor& filter_mats = filter_mat.filter_mats_;
+        const auto f_size_d4 = filter_mat.filter_shape_.size_dim_4_;
+        const auto f_height = filter_mat.filter_shape_.height_;
+        const auto f_width = filter_mat.filter_shape_.width_;
+        const auto f_depth = filter_mat.filter_shape_.depth_;
+        const auto out_depth = filter_mat.filter_count_;
+
+        assertion(f_depth == in.shape().depth_, "filter depth does not match input");
+        assertion(filter_mats.shape().size_dim_5_ == f_size_d4, "incorrect number of filter levels in d4 direction");
+        assertion(filter_mats.shape().size_dim_4_ == f_height, "incorrect number of filter levels in y direction");
+        assertion(out_width == (in.shape().width_ - f_width) / strides_x + 1, "output width does not match");
+        assertion(out_depth == filter_mat.biases_.size(), "invalid bias count");
+
+        tensor output = init_conv_output_tensor_3d(out_size_d4, out_height, out_width, out_depth, in.shape().rank(), filter_mat);
+
+        for (std::size_t d4_filt = 0; d4_filt < f_size_d4; ++d4_filt) {
+            for (std::size_t y_filt = 0; y_filt < f_height; ++y_filt) {
+                const Eigen::Map<ColMajorMatrixXf, Eigen::Unaligned>
+                    filter(const_cast<float_type*>(&filter_mats.get_ref_ignore_rank(tensor_pos(d4_filt, y_filt, 0, 0, 0))),
+                        static_cast<EigenIndex>(out_depth),
+                        static_cast<EigenIndex>(f_width * f_depth));
+                for (std::size_t d4 = 0, d4_out = 0; d4 < in.shape().size_dim_4_ + 1 - f_size_d4; d4 += strides_d4, ++d4_out) {
+                    for (std::size_t y = 0, y_out = 0; y < in.shape().height_ + 1 - f_height; y += strides_y, ++y_out) {
+                        const auto input = get_im2col_mapping_3d(in, f_width, f_depth, strides_x, out_width, d4, y, d4_filt, y_filt);
+                        Eigen::Map<ColMajorMatrixXf, Eigen::Unaligned>
+                            output_map(&output.get_ref_ignore_rank(tensor_pos(0, d4_out, y_out, 0, 0)),
+                                static_cast<EigenIndex>(out_depth),
+                                static_cast<EigenIndex>(out_width));
+
+                        output_map.noalias() += filter * input;
+                    }
+                }
+            }
+        }
+
+        return output;
+    }
+
+    inline tensor convolve_3d(
+        const shape3& strides,
+        const padding& pad_type,
+        const convolution3d_filter_matrices& filter_mat,
+        const tensor& input)
+    {
+        assertion(filter_mat.filter_shape_.depth_ == input.shape().depth_,
+            "invalid filter depth");
+
+        const shape3 filter_spatial_shape(
+            filter_mat.filter_shape_.size_dim_4_,
+            filter_mat.filter_shape_.height_,
+            filter_mat.filter_shape_.width_);
+
+        const auto conv_cfg = preprocess_convolution_3d(
+            filter_spatial_shape,
+            strides, pad_type,
+            input.shape().size_dim_4_,
+            input.shape().height_,
+            input.shape().width_);
+
+        const auto in_padded = pad_tensor(0,
+            conv_cfg.pad_front_, conv_cfg.pad_back_,
+            conv_cfg.pad_top_, conv_cfg.pad_bottom_,
+            conv_cfg.pad_left_, conv_cfg.pad_right_,
+            input);
+
+        return convolve_accumulative_3d(
+            conv_cfg.out_size_d4_, conv_cfg.out_height_, conv_cfg.out_width_,
+            strides.size_dim_4_, strides.height_, strides.width_,
+            filter_mat,
+            in_padded);
+    }
+
 }
 }
diff --git a/include/fdeep/import_model.hpp b/include/fdeep/import_model.hpp
index 2a2bafd0..0b46e393 100644
--- a/include/fdeep/import_model.hpp
+++ b/include/fdeep/import_model.hpp
@@ -39,6 +39,7 @@
 #include "fdeep/layers/concatenate_layer.hpp"
 #include "fdeep/layers/conv_2d_layer.hpp"
 #include "fdeep/layers/conv_2d_transpose_layer.hpp"
+#include "fdeep/layers/conv_3d_layer.hpp"
 #include "fdeep/layers/cropping_3d_layer.hpp"
 #include "fdeep/layers/dense_layer.hpp"
 #include "fdeep/layers/depthwise_conv_2d_layer.hpp"
@@ -416,6 +417,36 @@ namespace internal {
             dilation_rate, weights, bias);
     }
 
+    inline layer_ptr create_conv_3d_layer(const get_param_f& get_param,
+        const nlohmann::json& data,
+        const std::string& name)
+    {
+        const std::string padding_str = data["config"]["padding"];
+        const auto pad_type = create_padding(padding_str);
+
+        const shape3 strides = create_shape3(data["config"]["strides"]);
+        const shape3 dilation_rate = create_shape3(data["config"]["dilation_rate"]);
+
+        const auto filter_count = create_size_t(data["config"]["filters"]);
+        float_vec bias(filter_count, 0);
+        const bool use_bias = data["config"]["use_bias"];
+        if (use_bias)
+            bias = decode_floats(get_param(name, "bias"));
+        assertion(bias.size() == filter_count, "size of bias does not match");
+
+        const float_vec weights = decode_floats(get_param(name, "weights"));
+        const shape3 kernel_size = create_shape3(data["config"]["kernel_size"]);
+        assertion(weights.size() % kernel_size.volume() == 0,
+            "invalid number of weights");
+        const std::size_t filter_depths = weights.size() / (kernel_size.volume() * filter_count);
+        const tensor_shape filter_shape(
+            kernel_size.size_dim_4_, kernel_size.height_, kernel_size.width_, filter_depths);
+
+        return std::make_shared<conv_3d_layer>(name,
+            filter_shape, filter_count, strides, pad_type,
+            dilation_rate, weights, bias);
+    }
+
     inline layer_ptr create_conv_2d_transpose_layer(const get_param_f& get_param,
         const nlohmann::json& data,
         const std::string& name)
@@ -1267,6 +1298,7 @@ namespace internal {
             { "Identity", create_identity_layer },
             { "Conv1D", create_conv_2d_layer },
             { "Conv2D", create_conv_2d_layer },
+            { "Conv3D", create_conv_3d_layer },
             { "Conv1DTranspose", create_conv_2d_transpose_layer },
             { "Conv2DTranspose", create_conv_2d_transpose_layer },
             { "SeparableConv1D", create_separable_conv_2D_layer },
diff --git a/include/fdeep/layers/conv_3d_layer.hpp b/include/fdeep/layers/conv_3d_layer.hpp
new file mode 100644
index 00000000..d7df9bc6
--- /dev/null
+++ b/include/fdeep/layers/conv_3d_layer.hpp
@@ -0,0 +1,54 @@
+// Copyright 2016, Tobias Hermann.
+// https://github.com/Dobiasd/frugally-deep
+// Distributed under the MIT License.
+// (See accompanying LICENSE file or at
+//  https://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include "fdeep/convolution3d.hpp"
+#include "fdeep/filter.hpp"
+#include "fdeep/layers/layer.hpp"
+#include "fdeep/shape3.hpp"
+#include "fdeep/tensor_shape.hpp"
+
+#include <fplus/fplus.hpp>
+
+#include <cstddef>
+#include <string>
+#include <vector>
+
+namespace fdeep {
+namespace internal {
+
+    class conv_3d_layer : public layer {
+    public:
+        explicit conv_3d_layer(
+            const std::string& name, const tensor_shape& filter_shape,
+            std::size_t k, const shape3& strides, padding p,
+            const shape3& dilation_rate,
+            const float_vec& weights, const float_vec& bias)
+            : layer(name)
+            , filters_(generate_im2col_filter_matrix_3d(
+                  generate_filters_3d(dilation_rate, filter_shape, k, weights, bias)))
+            , strides_(strides)
+            , padding_(p)
+        {
+            assertion(k > 0, "needs at least one filter");
+            assertion(filter_shape.volume() > 0, "filter must have volume");
+            assertion(strides.volume() > 0, "invalid strides");
+        }
+
+    protected:
+        tensors apply_impl(const tensors& inputs) const override
+        {
+            const auto& input = single_tensor_from_tensors(inputs);
+            return { convolve_3d(strides_, padding_, filters_, input) };
+        }
+        convolution3d_filter_matrices filters_;
+        shape3 strides_;
+        padding padding_;
+    };
+
+}
+}
diff --git a/keras_export/convert_model.py b/keras_export/convert_model.py
index b1137cc8..3aa58fff 100755
--- a/keras_export/convert_model.py
+++ b/keras_export/convert_model.py
@@ -222,6 +222,12 @@ def prepare_filter_weights_conv_2d_transpose(weights: NDFloat32Array) -> NDFloat
     return np.moveaxis(weights, [0, 1, 2, 3], [1, 2, 0, 3]).flatten()
 
 
+def prepare_filter_weights_conv_3d(weights: NDFloat32Array) -> NDFloat32Array:
+    """Change dimension order of 3d filter weights to the one used in fdeep"""
+    assert len(weights.shape) == 5
+    return np.moveaxis(weights, [0, 1, 2, 3, 4], [1, 2, 3, 4, 0]).flatten()
+
+
 def show_conv_1d_layer(layer: Layer) -> Mapping[str, list[str]]:
     """Serialize Conv1D layer to dict"""
     weights = layer.get_weights()
@@ -260,6 +266,25 @@ def show_conv_2d_layer(layer: Layer) -> Mapping[str, list[str]]:
     return result
 
 
+def show_conv_3d_layer(layer: Layer) -> Mapping[str, list[str]]:
+    """Serialize Conv3D layer to dict"""
+    weights = layer.get_weights()
+    assert len(weights) == 1 or len(weights) == 2
+    assert len(weights[0].shape) == 5
+    weights_flat = prepare_filter_weights_conv_3d(weights[0])
+    assert layer.padding in ['valid', 'same']
+    assert layer.groups == 1
+    assert len(get_layer_input_shape(layer)) == 5
+    assert get_layer_input_shape(layer)[0] in {None, 1}
+    result = {
+        'weights': encode_floats(weights_flat)
+    }
+    if len(weights) == 2:
+        bias = weights[1]
+        result['bias'] = encode_floats(bias)
+    return result
+
+
 def show_separable_conv_2d_layer(layer: Layer) -> Mapping[str, list[str]]:
     """Serialize SeparableConv2D layer to dict"""
     weights = layer.get_weights()
@@ -498,6 +523,7 @@ def get_layer_functions_dict() -> Mapping[str, Callable[[Layer], LayerConfig]]:
     return {
         'Conv1D': show_conv_1d_layer,
         'Conv2D': show_conv_2d_layer,
+        'Conv3D': show_conv_3d_layer,
         'Conv1DTranspose': show_conv_1d_transpose_layer,
         'Conv2DTranspose': show_conv_2d_transpose_layer,
         'SeparableConv2D': show_separable_conv_2d_layer,
diff --git a/keras_export/generate_test_models.py b/keras_export/generate_test_models.py
index dc2da305..b2b9522c 100644
--- a/keras_export/generate_test_models.py
+++ b/keras_export/generate_test_models.py
@@ -14,6 +14,7 @@
 from keras.layers import CategoryEncoding, Embedding
 from keras.layers import Conv1D, ZeroPadding1D, Cropping1D
 from keras.layers import Conv2D, ZeroPadding2D, Cropping2D, CenterCrop
+from keras.layers import Conv3D
 from keras.layers import GlobalAveragePooling1D, GlobalMaxPooling1D
 from keras.layers import GlobalAveragePooling2D, GlobalMaxPooling2D
 from keras.layers import GlobalAveragePooling3D, GlobalMaxPooling3D
@@ -230,6 +231,11 @@ def get_test_model_exhaustive() -> Model:
     outputs.append(Conv2D(4, (2, 4), strides=(2, 3), padding='same')(inputs[4]))
     outputs.append(Conv2D(4, (2, 4), padding='same', dilation_rate=(2, 3))(inputs[4]))
 
+    outputs.append(Conv3D(4, (2, 3, 3))(inputs[2]))
+    outputs.append(Conv3D(4, (2, 3, 3), use_bias=False, padding='valid')(inputs[2]))
+    outputs.append(Conv3D(4, (1, 2, 4), strides=(2, 2, 3), padding='same')(inputs[2]))
+    outputs.append(Conv3D(4, (1, 2, 4), padding='same', dilation_rate=(2, 2, 3))(inputs[2]))
+
     outputs.append(SeparableConv2D(3, (3, 3))(inputs[4]))
     outputs.append(DepthwiseConv2D((3, 3))(inputs[4]))
     outputs.append(DepthwiseConv2D((1, 2))(inputs[4]))