Added skeletal animations and other fixes

2026-03-16 12:57:53 +01:00
parent 2cc18b3329
commit 28c6703892
51 changed files with 1964 additions and 602 deletions
--- a/.gitmodules
+++ b/.gitmodules
@@ -38,3 +38,6 @@
 [submodule "destrum/third_party/tinyexr"]
 	path = destrum/third_party/tinyexr
 	url = https://github.com/syoyo/tinyexr.git
+[submodule "destrum/third_party/assimp"]
+	path = destrum/third_party/assimp
+	url = https://github.com/assimp/assimp.git
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -22,4 +22,4 @@ add_custom_target(CookAssets)
 add_dependencies(CookAssets
        _internal_cook_game_assets
        _internal_cook_engine_assets
-)
+)
--- a/2
+++ b/2
--- a/destrum/CMakeLists.txt
+++ b/destrum/CMakeLists.txt
@@ -4,6 +4,7 @@ set(SRC_FILES
        "src/App.cpp"
        "src/Event.cpp"

+        "src/Components/Animator.cpp"
        "src/Components/MeshRendererComponent.cpp"
        "src/Components/Rotator.cpp"
        "src/Components/Spinner.cpp"
@@ -11,6 +12,8 @@ set(SRC_FILES

        "src/Graphics/BindlessSetManager.cpp"
        "src/Graphics/Camera.cpp"
+        "src/Graphics/ComputePipeline.cpp"
+        "src/Graphics/Frustum.cpp"
        "src/Graphics/GfxDevice.cpp"
        "src/Graphics/ImageCache.cpp"
        "src/Graphics/ImageLoader.cpp"
@@ -29,6 +32,7 @@ set(SRC_FILES

        "src/Graphics/Pipelines/MeshPipeline.cpp"
        "src/Graphics/Pipelines/SkyboxPipeline.cpp"
+        "src/Graphics/Pipelines/SkinningPipeline.cpp"

        "src/Input/InputManager.cpp"

@@ -69,6 +73,7 @@ target_link_libraries(destrum
        spdlog::spdlog
        stb::image
        tinygltf
+        assimp

        PRIVATE
        freetype::freetype
@@ -105,14 +110,14 @@ target_compile_definitions(destrum

 set(ASSETS_SRC_DIR     "${CMAKE_CURRENT_LIST_DIR}/assets_src")
 set(ASSETS_RUNTIME_DIR "${CMAKE_CURRENT_LIST_DIR}/assets_runtime")
-set(OUTPUT_ENGINE_ASSETS_DIR "${CMAKE_BINARY_DIR}/assets/engine")
-
-add_custom_command(TARGET destrum POST_BUILD
-        COMMAND ${CMAKE_COMMAND} -E rm -rf "${OUTPUT_ENGINE_ASSETS_DIR}"
-        COMMAND ${CMAKE_COMMAND} -E make_directory "${CMAKE_BINARY_DIR}/assets"
-        COMMAND ${CMAKE_COMMAND} -E create_symlink "${ASSETS_RUNTIME_DIR}" "${OUTPUT_ENGINE_ASSETS_DIR}"
-        VERBATIM
-)
+#set(OUTPUT_ENGINE_ASSETS_DIR "${CMAKE_BINARY_DIR}/assets/engine")
+#
+#add_custom_command(TARGET destrum POST_BUILD
+#        COMMAND ${CMAKE_COMMAND} -E rm -rf "${OUTPUT_ENGINE_ASSETS_DIR}"
+#        COMMAND ${CMAKE_COMMAND} -E make_directory "${CMAKE_BINARY_DIR}/assets"
+#        COMMAND ${CMAKE_COMMAND} -E create_symlink "${ASSETS_RUNTIME_DIR}" "${OUTPUT_ENGINE_ASSETS_DIR}"
+#        VERBATIM
+#)

 add_custom_target(_internal_clean_engine_assets
        COMMAND TheChef
@@ -127,3 +132,17 @@ add_custom_target(_internal_cook_engine_assets ALL
        --output "${ASSETS_RUNTIME_DIR}"
        DEPENDS TheChef
 )
+
+function(destrum_cook_engine_assets GAME_TARGET GAME_OUTPUT_DIR)
+    # This resolves to destrum's own directory at function DEFINITION time
+    set(_engine_src     "${CMAKE_CURRENT_FUNCTION_LIST_DIR}/assets_src")
+    set(_engine_runtime "${CMAKE_CURRENT_FUNCTION_LIST_DIR}/assets_runtime")
+    set(_output_engine  "${GAME_OUTPUT_DIR}/assets/engine")
+
+    add_custom_command(TARGET ${GAME_TARGET} POST_BUILD
+            COMMAND ${CMAKE_COMMAND} -E make_directory "${GAME_OUTPUT_DIR}/assets"
+            COMMAND ${CMAKE_COMMAND} -E rm -rf "${_output_engine}"
+            COMMAND ${CMAKE_COMMAND} -E create_symlink "${_engine_runtime}" "${_output_engine}"
+            VERBATIM
+    )
+endfunction()
--- a/destrum/assets_src/char.fbx
+++ b/destrum/assets_src/char.fbx
--- a/destrum/assets_src/char.jpg
+++ b/destrum/assets_src/char.jpg
--- a/destrum/assets_src/char2.fbx
+++ b/destrum/assets_src/char2.fbx
--- a/destrum/assets_src/shaders/bindless.glsl
+++ b/destrum/assets_src/shaders/bindless.glsl
@@ -14,6 +14,10 @@ vec4 sampleTexture2DNearest(uint texID, vec2 uv) {
    return texture(nonuniformEXT(sampler2D(textures[texID], samplers[NEAREST_SAMPLER_ID])), uv);
 }

+vec4 sampleTexture2D(uint texID, uint samplerID, vec2 uv) {
+    return texture(nonuniformEXT(sampler2D(textures[texID], samplers[samplerID])), uv);
+}
+
 vec4 sampleTexture2DMSNearest(uint texID, ivec2 p, int s) {
    return texelFetch(nonuniformEXT(sampler2DMS(texturesMS[texID], samplers[NEAREST_SAMPLER_ID])), p, s);
 }
--- a/destrum/assets_src/shaders/materials.glsl
+++ b/destrum/assets_src/shaders/materials.glsl
@@ -6,6 +6,7 @@
 struct MaterialData {
    vec4 baseColor;
    vec4 metallicRoughnessEmissive;
+    uint textureFilteringMode;
    uint diffuseTex;
    uint normalTex;
    uint metallicRoughnessTex;
--- a/destrum/assets_src/shaders/mesh.frag
+++ b/destrum/assets_src/shaders/mesh.frag
@@ -17,8 +17,9 @@ layout (location = 0) out vec4 outFragColor;
 void main()
 {
    MaterialData material = pcs.sceneData.materials.data[pcs.materialID];
+    uint samplerID = pcs.sceneData.materials.data[pcs.materialID].textureFilteringMode;

-    vec4 diffuse = sampleTexture2DLinear(material.diffuseTex, inUV) * material.baseColor;
+    vec4 diffuse = sampleTexture2D(material.diffuseTex, samplerID, inUV) * material.baseColor;
    outFragColor = diffuse;


--- a/destrum/assets_src/shaders/skinning.comp
+++ b/destrum/assets_src/shaders/skinning.comp
@@ -0,0 +1,60 @@
+#version 460
+
+#extension GL_GOOGLE_include_directive : require
+#extension GL_EXT_buffer_reference : require
+
+#include "vertex.glsl"
+
+struct SkinningDataType {
+    ivec4 jointIds;
+    vec4 weights;
+};
+
+layout (buffer_reference, std430) readonly buffer SkinningData {
+	SkinningDataType data[];
+};
+
+layout (buffer_reference, std430) readonly buffer JointMatrices {
+	mat4 matrices[];
+};
+
+layout (push_constant) uniform constants
+{
+    JointMatrices jointMatrices;
+    uint jointMatricesStartIndex;
+    uint numVertices;
+	VertexBuffer inputBuffer;
+    SkinningData skinningData;
+	VertexBuffer outputBuffer;
+} pcs;
+
+layout (local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+mat4 getJointMatrix(int jointId) {
+    if (jointId < 0) return mat4(1.0);
+    return pcs.jointMatrices.matrices[pcs.jointMatricesStartIndex + jointId];
+}
+
+void main()
+{
+    uint index = gl_GlobalInvocationID.x;
+    if (index >= pcs.numVertices) {
+        return;
+    }
+
+    SkinningDataType sd = pcs.skinningData.data[index];
+    mat4 skinMatrix =
+        sd.weights.x * getJointMatrix(sd.jointIds.x) +
+        sd.weights.y * getJointMatrix(sd.jointIds.y) +
+        sd.weights.z * getJointMatrix(sd.jointIds.z) +
+        sd.weights.w * getJointMatrix(sd.jointIds.w);
+
+    Vertex v = pcs.inputBuffer.vertices[index];
+    v.position = vec3(skinMatrix * vec4(v.position, 1.0));
+
+    mat3 skinMat3 = mat3(skinMatrix);
+    v.normal = skinMat3 * v.normal;
+    v.tangent.xyz = skinMat3 * v.tangent.xyz; // don't transform tangent.w
+
+    pcs.outputBuffer.vertices[index] = v;
+}
--- a/destrum/include/destrum/Components/Animator.h
+++ b/destrum/include/destrum/Components/Animator.h
@@ -0,0 +1,70 @@
+#ifndef ANIMATOR_H
+#define ANIMATOR_H
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include <glm/mat4x4.hpp>
+
+#include <destrum/ObjectModel/Component.h>
+#include <destrum/Graphics/Resources/Mesh.h>
+#include <destrum/Graphics/SkeletalAnimation.h>
+
+class SkinningPipeline;
+
+class Animator : public Component {
+public:
+    explicit Animator(GameObject& parent);
+
+    // Component interface
+    void Update() override;
+    void ImGuiInspector() override;
+
+    // Animation control
+    void addClip(std::shared_ptr<SkeletalAnimation> clip);
+    void play(const std::string& name, float blendTime = 0.f);
+    void stop();
+
+    bool isPlaying() const { return m_current.clip != nullptr; }
+    const std::string& currentClipName() const { return m_currentClipName; }
+    float currentTime() const { return m_current.time; }
+
+    // Called during draw command building
+    std::size_t uploadJointMatrices(SkinningPipeline& pipeline,
+                                    const Skeleton& skeleton,
+                                    std::size_t frameIndex);
+    Skeleton* getSkeleton() {
+        return &m_skeleton;
+    }
+
+    void setSkeleton(Skeleton skeleton) {
+        m_skeleton = std::move(skeleton);
+    }
+
+private:
+    struct PlaybackState {
+        SkeletalAnimation* clip  = nullptr;
+        float              time  = 0.f;
+        float              speed = 1.f;
+    };
+
+    Skeleton m_skeleton{};
+
+    PlaybackState m_current;
+    PlaybackState m_previous;
+    float         m_blendT        = 0.f;
+    float         m_blendDuration = 0.f;
+    std::string   m_currentClipName;
+
+    std::unordered_map<std::string, std::shared_ptr<SkeletalAnimation>> m_clips;
+
+    std::vector<glm::mat4> computeJointMatrices(const Skeleton& skeleton);
+
+    glm::vec3 sampleTranslation(const SkeletalAnimation::Track& track, float t);
+    glm::quat sampleRotation   (const SkeletalAnimation::Track& track, float t);
+    glm::vec3 sampleScale      (const SkeletalAnimation::Track& track, float t);
+};
+
+#endif // ANIMATOR_H
--- a/destrum/include/destrum/Components/MeshRendererComponent.h
+++ b/destrum/include/destrum/Components/MeshRendererComponent.h
@@ -21,6 +21,8 @@ public:
 private:
    MeshID meshID{NULL_MESH_ID};
    MaterialID materialID{NULL_MATERIAL_ID};
+
+    std::unique_ptr<SkinnedMesh> m_skinnedMesh;
 };

 #endif //MESHRENDERERCOMPONENT_H
--- a/destrum/include/destrum/Graphics/Camera.h
+++ b/destrum/include/destrum/Graphics/Camera.h
@@ -96,6 +96,13 @@ public:
        this->CalculateProjectionMatrix();
    }

+    float GetZNear() const { return m_zNear; }
+    void SetZNear(float zNear) { m_zNear = zNear; }
+    float GetZFar() const { return m_zFar; }
+    void SetZFar(float zFar) { m_zFar = zFar; }
+    float GetFOVY() const { return fovAngle; }
+    float GetAspectRatio() const { return m_aspectRatio; }
+
    glm::vec3 m_position{2.f, 0, 0};

    void Translate(const glm::vec3& translation) {
--- a/destrum/include/destrum/Graphics/ComputePipeline.h
+++ b/destrum/include/destrum/Graphics/ComputePipeline.h
@@ -1,4 +1,48 @@
 #ifndef COMPUTEPIPELINE_H
 #define COMPUTEPIPELINE_H

+#include <string>
+#include <vector>
+
+#include <destrum/Graphics/GfxDevice.h>
+
+struct ComputePipelineConfigInfo {
+    std::string name;
+
+    // Optional specialization constants (can be nullptr if unused)
+    VkSpecializationInfo* specializationInfo{nullptr};
+
+    VkPipelineLayout pipelineLayout{VK_NULL_HANDLE};
+};
+
+class ComputePipeline {
+public:
+    ComputePipeline(GfxDevice& device,
+                    const std::string& compPath,
+                    const ComputePipelineConfigInfo& configInfo);
+    ~ComputePipeline();
+
+    ComputePipeline(const ComputePipeline& other) = delete;
+    ComputePipeline(ComputePipeline&& other) noexcept = delete;
+    ComputePipeline& operator=(const ComputePipeline& other) = delete;
+    ComputePipeline& operator=(ComputePipeline&& other) noexcept = delete;
+
+    void bind(VkCommandBuffer buffer) const;
+
+    static void DefaultPipelineConfigInfo(ComputePipelineConfigInfo& configInfo);
+
+private:
+    static std::vector<char> readFile(const std::string& filename);
+
+    void CreateComputePipeline(const std::string& compPath,
+                               const ComputePipelineConfigInfo& configInfo);
+
+    void CreateShaderModule(const std::vector<char>& code, VkShaderModule* shaderModule) const;
+
+    GfxDevice& m_device;
+
+    VkPipeline m_computePipeline{VK_NULL_HANDLE};
+    VkShaderModule m_compShaderModule{VK_NULL_HANDLE};
+};
+
 #endif //COMPUTEPIPELINE_H
--- a/destrum/include/destrum/Graphics/Frustum.h
+++ b/destrum/include/destrum/Graphics/Frustum.h
@@ -1,4 +1,80 @@
 #ifndef FRUSTUM_H
 #define FRUSTUM_H

+#include <glm/glm.hpp>
+
+#include "Camera.h"
+
+struct Frustum {
+    struct Plane {
+        Plane() = default;
+        Plane(const glm::vec3& p1, const glm::vec3& norm) :
+            normal(glm::normalize(norm)), distance(glm::dot(normal, p1))
+        {}
+
+        glm::vec3 normal{0.f, 1.f, 0.f};
+
+        // distance from the origin to the nearest point in the plane
+        float distance{0.f};
+
+        float getSignedDistanceToPlane(const glm::vec3& point) const
+        {
+            return glm::dot(normal, point) - distance;
+        }
+    };
+
+    const Plane& getPlane(int i) const
+    {
+        switch (i) {
+            case 0:
+                return farFace;
+            case 1:
+                return nearFace;
+            case 2:
+                return leftFace;
+            case 3:
+                return rightFace;
+            case 4:
+                return topFace;
+            case 5:
+                return bottomFace;
+            default:
+                assert(false);
+                return nearFace;
+        }
+    }
+
+    Plane farFace;
+    Plane nearFace;
+
+    Plane leftFace;
+    Plane rightFace;
+
+    Plane topFace;
+    Plane bottomFace;
+};
+
+struct Sphere {
+    glm::vec3 center{};
+    float radius{};
+};
+
+struct AABB {
+    glm::vec3 min;
+    glm::vec3 max;
+
+    glm::vec3 calculateSize() const { return glm::abs(max - min); }
+};
+
+namespace edge
+{
+    std::array<glm::vec3, 8> calculateFrustumCornersWorldSpace(const Camera& camera);
+
+    Frustum createFrustumFromCamera(const Camera& camera);
+    bool isInFrustum(const Frustum& frustum, const Sphere& s);
+    bool isInFrustum(const Frustum& frustum, const AABB& aabb);
+    Sphere calculateBoundingSphereWorld(const glm::mat4& transform, const Sphere& s, bool hasSkeleton);
+}
+
+
 #endif //FRUSTUM_H
--- a/destrum/include/destrum/Graphics/GfxDevice.h
+++ b/destrum/include/destrum/Graphics/GfxDevice.h
@@ -26,6 +26,8 @@

 #include "Util.h"

+class MeshCache;
+
 namespace {
    using ImmediateExecuteFunction = std::function<void(VkCommandBuffer)>;
 }
--- a/destrum/include/destrum/Graphics/Material.h
+++ b/destrum/include/destrum/Graphics/Material.h
@@ -4,13 +4,21 @@
 #include <string>
 #include <glm/glm.hpp>

-struct MaterialData {
-    glm::vec4 baseColor;
-    glm::vec4 metalRoughnessEmissive;
-    std::uint32_t diffuseTex;
-    std::uint32_t normalTex;
-    std::uint32_t metallicRoughnessTex;
-    std::uint32_t emissiveTex;
+struct alignas(16) MaterialData {
+    alignas(16) glm::vec4 baseColor;
+    alignas(16) glm::vec4 metalRoughnessEmissive;
+    uint32_t textureFilteringMode;
+    uint32_t diffuseTex;
+    uint32_t normalTex;
+    uint32_t metallicRoughnessTex;
+    uint32_t emissiveTex;
+    uint32_t _pad0, _pad1, _pad2; // explicit to 64 bytes
+};
+
+enum class TextureFilteringMode : std::uint32_t {
+    Nearest,
+    Linear,
+    Anisotropic,
 };

 struct Material {
@@ -19,6 +27,8 @@ struct Material {
    float roughnessFactor{0.7f};
    float emissiveFactor{0.f};

+    TextureFilteringMode textureFilteringMode{TextureFilteringMode::Linear};
+
    ImageID diffuseTexture{NULL_IMAGE_ID};
    // ImageId normalMapTexture{NULL_IMAGE_ID};
    // ImageId metallicRoughnessTexture{NULL_IMAGE_ID};
--- a/destrum/include/destrum/Graphics/MaterialCache.h
+++ b/destrum/include/destrum/Graphics/MaterialCache.h
@@ -18,13 +18,14 @@ public:
    void cleanup(GfxDevice& gfxDevice);

    MaterialID addMaterial(GfxDevice& gfxDevice, Material material);
-    const Material& getMaterial(MaterialID id) const;
+    MaterialID addSimpleTextureMaterial(GfxDevice& gfxDevice, ImageID textureID);
+    [[nodiscard]] const Material& getMaterial(MaterialID id) const;

-    MaterialID getFreeMaterialId() const;
-    MaterialID getPlaceholderMaterialId() const;
+    [[nodiscard]] MaterialID getFreeMaterialId() const;
+    [[nodiscard]] MaterialID getPlaceholderMaterialId() const;

-    const GPUBuffer& getMaterialDataBuffer() const { return materialDataBuffer; }
-    VkDeviceAddress getMaterialDataBufferAddress() const { return materialDataBuffer.address; }
+    [[nodiscard]] const GPUBuffer& getMaterialDataBuffer() const { return materialDataBuffer; }
+    [[nodiscard]] VkDeviceAddress getMaterialDataBufferAddress() const { return materialDataBuffer.address; }


    Material& getMaterialMutable(MaterialID id);
@@ -33,7 +34,7 @@ public:
 private:
    std::vector<Material> materials;

-    static const auto MAX_MATERIALS = 1000;
+    static constexpr auto MAX_MATERIALS = 1000;
    GPUBuffer materialDataBuffer;

    // material which is used for meshes without materials
--- a/destrum/include/destrum/Graphics/MeshCache.h
+++ b/destrum/include/destrum/Graphics/MeshCache.h
@@ -13,11 +13,13 @@ public:

    MeshID addMesh(GfxDevice& gfxDevice, const CPUMesh& cpuMesh);
    const GPUMesh& getMesh(MeshID id) const;
+    const CPUMesh& getCPUMesh(MeshID id) const;

 private:
    void uploadMesh(GfxDevice& gfxDevice, const CPUMesh& cpuMesh, GPUMesh& gpuMesh) const;

    std::vector<GPUMesh> meshes;
+    std::vector<CPUMesh> cpuMeshes;
 };

 #endif //MESHCACHE_H
--- a/destrum/include/destrum/Graphics/MeshDrawCommand.h
+++ b/destrum/include/destrum/Graphics/MeshDrawCommand.h
@@ -5,13 +5,15 @@

 #include <destrum/Graphics/ids.h>

+#include <destrum/Graphics/Frustum.h>
+

 struct MeshDrawCommand {
    MeshID meshId;
    glm::mat4 transformMatrix;

    // for frustum culling
-    // math::Sphere worldBoundingSphere;
+    Sphere worldBoundingSphere;

    // If set - mesh will be drawn with overrideMaterialId
    // instead of whatever material the mesh has
@@ -20,8 +22,8 @@ struct MeshDrawCommand {
    bool castShadow{true};

    // skinned meshes only
-    // const SkinnedMesh* skinnedMesh{nullptr};
-    // std::uint32_t jointMatricesStartIndex;
+    const SkinnedMesh* skinnedMesh{nullptr};
+    std::uint32_t jointMatricesStartIndex;
 };

 #endif //MESHDRAWCOMMAND_H
--- a/destrum/include/destrum/Graphics/Pipelines/SkinningPipeline.h
+++ b/destrum/include/destrum/Graphics/Pipelines/SkinningPipeline.h
@@ -1,4 +1,58 @@
 #ifndef SKINNINGPIPELINE_H
 #define SKINNINGPIPELINE_H

+#include <vulkan/vulkan.h>
+#include <array>
+#include <span>
+#include <glm/glm.hpp>
+
+#include <destrum/Graphics/ComputePipeline.h>
+#include <destrum/Graphics/Resources/AppendableBuffer.h>
+
+#include "destrum/Graphics/Resources/NBuffer.h"
+
+struct MeshDrawCommand;
+class MeshCache;
+class GfxDevice;
+
+class SkinningPipeline final {
+public:
+    void init(GfxDevice& gfxDevice);
+    void cleanup(GfxDevice& gfxDevice);
+
+    void doSkinning(
+        VkCommandBuffer cmd,
+        std::size_t frameIndex,
+        const MeshCache& meshCache,
+        const MeshDrawCommand& dc);
+
+    void beginDrawing(std::size_t frameIndex);
+    std::size_t appendJointMatrices(
+        std::span<const glm::mat4> jointMatrices,
+        std::size_t frameIndex);
+
+private:
+    VkPipelineLayout m_pipelineLayout;
+    std::unique_ptr<ComputePipeline> skinningPipeline;
+    struct PushConstants {
+        VkDeviceAddress jointMatricesBuffer;
+        std::uint32_t jointMatricesStartIndex;
+        std::uint32_t numVertices;
+        VkDeviceAddress inputBuffer;
+        VkDeviceAddress skinningData;
+        VkDeviceAddress outputBuffer;
+    };
+    static constexpr std::size_t MAX_JOINT_MATRICES = 5000;
+
+    struct PerFrameData {
+        AppendableBuffer<glm::mat4> jointMatricesBuffer;
+    };
+
+    std::array<PerFrameData, FRAMES_IN_FLIGHT> framesData;
+    // NBuffer framesData;
+
+    PerFrameData& getCurrentFrameData(std::size_t frameIndex);
+};
+
+
 #endif //SKINNINGPIPELINE_H
--- a/destrum/include/destrum/Graphics/Renderer.h
+++ b/destrum/include/destrum/Graphics/Renderer.h
@@ -13,6 +13,7 @@

 #include <memory>

+#include <destrum/Graphics/Pipelines/SkinningPipeline.h>
 #include "Pipelines/SkyboxPipeline.h"

 class GameRenderer {
@@ -35,6 +36,11 @@ public:
    void cleanup(VkDevice device);

    void drawMesh(MeshID id, const glm::mat4& transform, MaterialID materialId);
+    void drawSkinnedMesh(MeshID id,
+                     const glm::mat4& transform,
+                     MaterialID materialId,
+                     SkinnedMesh* skinnedMesh,
+                     std::size_t jointMatricesStartIndex);
    const GPUImage& getDrawImage(const GfxDevice& gfx_device) const;

    void resize(GfxDevice& gfxDevice, const glm::ivec2& newSize) {
@@ -47,6 +53,13 @@ public:
    void setSkyboxTexture(ImageID skyboxImageId);

    void flushMaterialUpdates(GfxDevice& gfxDevice);
+    [[nodiscard]] MeshCache& GetMeshCache() const {
+        return meshCache;
+    }
+
+    [[nodiscard]] SkinningPipeline& getSkinningPipeline() const {
+        return *skinningPipeline;
+    }

 private:
    void createDrawImage(GfxDevice& gfxDevice, const glm::ivec2& drawImageSize, bool firstCreate);
@@ -91,6 +104,8 @@ private:

    std::unique_ptr<MeshPipeline> meshPipeline;
    std::unique_ptr<SkyboxPipeline> skyboxPipeline;
+
+    std::unique_ptr<SkinningPipeline> skinningPipeline;
 };

 #endif //RENDERER_H
--- a/destrum/include/destrum/Graphics/Resources/AnimationClip.h
+++ b/destrum/include/destrum/Graphics/Resources/AnimationClip.h
@@ -0,0 +1,30 @@
+#ifndef ANIMATIONCLIP_H
+#define ANIMATIONCLIP_H
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include <glm/glm.hpp>
+
+struct JointKeyframes {
+    std::uint32_t jointIndex;  // matches your skeleton joint order
+
+    std::vector<float>      positionTimes;
+    std::vector<glm::vec3>  positions;
+
+    std::vector<float>      rotationTimes;
+    std::vector<glm::quat>  rotations;
+
+    std::vector<float>      scaleTimes;
+    std::vector<glm::vec3>  scales;
+};
+
+struct AnimationClip {
+    std::string name;
+    float       duration;        // seconds
+    float       ticksPerSecond;
+    std::vector<JointKeyframes> channels;  // one per animated joint
+};
+
+#endif //ANIMATIONCLIP_H
--- a/destrum/include/destrum/Graphics/Resources/AppendableBuffer.h
+++ b/destrum/include/destrum/Graphics/Resources/AppendableBuffer.h
@@ -1,4 +1,40 @@
 #ifndef APPENDABLEBUFFER_H
 #define APPENDABLEBUFFER_H

+#include <cassert>
+#include <cstring> // memcpy
+#include <span>
+
+#include <vulkan/vulkan.h>
+
+#include <destrum/Graphics/Resources/Buffer.h>
+
+template<typename T>
+struct AppendableBuffer {
+    void append(std::span<const T> elements)
+    {
+        assert(size + elements.size() <= capacity);
+        auto arr = (T*)buffer.info.pMappedData;
+        std::memcpy((void*)&arr[size], elements.data(), elements.size() * sizeof(T));
+        size += elements.size();
+    }
+
+    void append(const T& element)
+    {
+        assert(size + 1 <= capacity);
+        auto arr = (T*)buffer.info.pMappedData;
+        std::memcpy((void*)&arr[size], &element, sizeof(T));
+        ++size;
+    }
+
+    void clear() { size = 0; }
+
+    VkBuffer getVkBuffer() const { return buffer.buffer; }
+
+    GPUBuffer buffer;
+    std::size_t capacity{};
+    std::size_t size{0};
+};
+
+
 #endif //APPENDABLEBUFFER_H
--- a/destrum/include/destrum/Graphics/Resources/Mesh.h
+++ b/destrum/include/destrum/Graphics/Resources/Mesh.h
@@ -2,33 +2,48 @@
 #define MESH_H

 #include <cstdint>
+#include <optional>
 #include <string>
+#include <unordered_map>
 #include <vector>

 #include <glm/vec2.hpp>
 #include <glm/vec3.hpp>
 #include <glm/vec4.hpp>
+#include <glm/mat4x4.hpp>

 #include <destrum/Graphics/Resources/Buffer.h>
+#include <destrum/Graphics/Frustum.h>
+#include <destrum/Graphics/Skeleton.h>

+// ─── CPU Mesh ─────────────────────────────────────────────────────────────────
 struct CPUMesh {
-    std::vector<std::uint32_t> indices;
-
    struct Vertex {
        glm::vec3 position;
-        float uv_x{};
+        float     uv_x{};
        glm::vec3 normal;
-        float uv_y{};
+        float     uv_y{};
        glm::vec4 tangent;
    };
-    std::vector<Vertex> vertices;
+
+    struct SkinningData {
+        glm::vec<4, std::uint32_t> jointIds;
+        glm::vec4                  weights;
+    };
+
+    std::vector<std::uint32_t> indices;
+    std::vector<Vertex>        vertices;
+    std::vector<SkinningData>  skinningData; // empty if no skeleton

    std::string name;

    glm::vec3 minPos;
    glm::vec3 maxPos;
+
+    std::optional<Skeleton> skeleton; // present only for skinned meshes
 };

+// ─── GPU Mesh ─────────────────────────────────────────────────────────────────

 struct GPUMesh {
    GPUBuffer vertexBuffer;
@@ -40,78 +55,64 @@ struct GPUMesh {
    // AABB
    glm::vec3 minPos;
    glm::vec3 maxPos;
-    // math::Sphere boundingSphere;
+    Sphere    boundingSphere;
+
+    bool      hasSkeleton{false};
+    GPUBuffer skinningDataBuffer; // valid only when hasSkeleton == true
 };

-static std::vector<CPUMesh::Vertex> vertices = {
-    // =======================
+// ─── Skinned output buffer (one per entity, not per mesh asset) ───────────────
+
+struct SkinnedMesh {
+    GPUBuffer skinnedVertexBuffer;
+};
+
+// ─── Cube geometry (dev / test asset) ─────────────────────────────────────────
+
+namespace CubeMesh {
+
+inline std::vector<CPUMesh::Vertex> vertices = {
    // +Z (Front)
-    // =======================
    {{-0.5f, -0.5f,  0.5f}, 0.0f, {0, 0, 1}, 0.0f, {1, 0, 0, 1}},
    {{ 0.5f, -0.5f,  0.5f}, 1.0f, {0, 0, 1}, 0.0f, {1, 0, 0, 1}},
    {{ 0.5f,  0.5f,  0.5f}, 1.0f, {0, 0, 1}, 1.0f, {1, 0, 0, 1}},
    {{-0.5f,  0.5f,  0.5f}, 0.0f, {0, 0, 1}, 1.0f, {1, 0, 0, 1}},
-
-    // =======================
    // -Z (Back)
-    // =======================
    {{ 0.5f, -0.5f, -0.5f}, 0.0f, {0, 0, -1}, 0.0f, {-1, 0, 0, 1}},
    {{-0.5f, -0.5f, -0.5f}, 1.0f, {0, 0, -1}, 0.0f, {-1, 0, 0, 1}},
    {{-0.5f,  0.5f, -0.5f}, 1.0f, {0, 0, -1}, 1.0f, {-1, 0, 0, 1}},
    {{ 0.5f,  0.5f, -0.5f}, 0.0f, {0, 0, -1}, 1.0f, {-1, 0, 0, 1}},
-
-    // =======================
    // +X (Right)
-    // =======================
    {{ 0.5f, -0.5f,  0.5f}, 0.0f, {1, 0, 0}, 0.0f, {0, 0, -1, 1}},
    {{ 0.5f, -0.5f, -0.5f}, 1.0f, {1, 0, 0}, 0.0f, {0, 0, -1, 1}},
    {{ 0.5f,  0.5f, -0.5f}, 1.0f, {1, 0, 0}, 1.0f, {0, 0, -1, 1}},
    {{ 0.5f,  0.5f,  0.5f}, 0.0f, {1, 0, 0}, 1.0f, {0, 0, -1, 1}},
-
-    // =======================
    // -X (Left)
-    // =======================
    {{-0.5f, -0.5f, -0.5f}, 0.0f, {-1, 0, 0}, 0.0f, {0, 0, 1, 1}},
    {{-0.5f, -0.5f,  0.5f}, 1.0f, {-1, 0, 0}, 0.0f, {0, 0, 1, 1}},
    {{-0.5f,  0.5f,  0.5f}, 1.0f, {-1, 0, 0}, 1.0f, {0, 0, 1, 1}},
    {{-0.5f,  0.5f, -0.5f}, 0.0f, {-1, 0, 0}, 1.0f, {0, 0, 1, 1}},
-
-    // =======================
    // +Y (Top)
-    // =======================
    {{-0.5f,  0.5f,  0.5f}, 0.0f, {0, 1, 0}, 0.0f, {1, 0, 0, 1}},
    {{ 0.5f,  0.5f,  0.5f}, 1.0f, {0, 1, 0}, 0.0f, {1, 0, 0, 1}},
    {{ 0.5f,  0.5f, -0.5f}, 1.0f, {0, 1, 0}, 1.0f, {1, 0, 0, 1}},
    {{-0.5f,  0.5f, -0.5f}, 0.0f, {0, 1, 0}, 1.0f, {1, 0, 0, 1}},
-
-    // =======================
    // -Y (Bottom)
-    // =======================
    {{-0.5f, -0.5f, -0.5f}, 0.0f, {0, -1, 0}, 0.0f, {1, 0, 0, 1}},
    {{ 0.5f, -0.5f, -0.5f}, 1.0f, {0, -1, 0}, 0.0f, {1, 0, 0, 1}},
    {{ 0.5f, -0.5f,  0.5f}, 1.0f, {0, -1, 0}, 1.0f, {1, 0, 0, 1}},
    {{-0.5f, -0.5f,  0.5f}, 0.0f, {0, -1, 0}, 1.0f, {1, 0, 0, 1}},
 };

-static std::vector<uint32_t> indices = {
-    // Front (+Z)
-    0, 2, 1,  0, 3, 2,
-
-    // Back (-Z)
-    4, 6, 5,  4, 7, 6,
-
-    // Right (+X)
-    8,10, 9,  8,11,10,
-
-    // Left (-X)
-   12,14,13, 12,15,14,
-
-    // Top (+Y)
-   16,18,17, 16,19,18,
-
-    // Bottom (-Y)
-   20,22,21, 20,23,22
+inline std::vector<std::uint32_t> indices = {
+     0, 2, 1,   0, 3, 2,  // Front
+     4, 6, 5,   4, 7, 6,  // Back
+     8,10, 9,   8,11,10,  // Right
+    12,14,13,  12,15,14,  // Left
+    16,18,17,  16,19,18,  // Top
+    20,22,21,  20,23,22,  // Bottom
 };

+} // namespace CubeMesh

-#endif //MESH_H
+#endif // MESH_H
--- a/destrum/include/destrum/Graphics/SkeletalAnimation.h
+++ b/destrum/include/destrum/Graphics/SkeletalAnimation.h
@@ -0,0 +1,35 @@
+#ifndef SKELETALANIMATION_H
+#define SKELETALANIMATION_H
+
+#include <map>
+#include <string>
+#include <vector>
+
+#include <glm/gtc/quaternion.hpp>
+#include <glm/vec3.hpp>
+
+struct SkeletalAnimation {
+    struct Keyframe {
+        float time;
+        glm::vec3 translation;
+        glm::quat rotation;
+        glm::vec3 scale;
+    };
+
+    struct Track {
+        std::uint32_t        jointIndex;
+        std::vector<Keyframe> keyframes; // sorted by time
+    };
+
+    std::vector<Track> tracks; // one per animated joint (not all joints need a track)
+    float              duration{0.f};
+    bool               looped{true};
+    std::string        name;
+
+    int                                      startFrame{0};
+    std::map<int, std::vector<std::string>>  events;
+
+    const std::vector<std::string>& getEventsForFrame(int frame) const;
+};
+
+#endif // SKELETALANIMATION_H
--- a/destrum/include/destrum/Graphics/Skeleton.h
+++ b/destrum/include/destrum/Graphics/Skeleton.h
@@ -0,0 +1,55 @@
+#ifndef SKELETON_H
+#define SKELETON_H
+
+#include <string>
+#include <vector>
+
+#include <glm/mat4x4.hpp>
+#include <glm/vec3.hpp>
+#include <glm/gtc/quaternion.hpp>
+
+#include <destrum/Graphics/ids.h>
+
+struct Joint {
+    JointId   id{NULL_JOINT_ID};
+    glm::vec3 localTranslation{0.f, 0.f, 0.f};
+    glm::quat localRotation{glm::identity<glm::quat>()};
+    glm::vec3 localScale{1.f, 1.f, 1.f};
+};
+
+struct Skeleton {
+    struct JointNode {
+        JointId id{NULL_JOINT_ID};
+        std::vector<JointId> children;
+    };
+
+    std::vector<JointNode>   hierarchy;
+    std::vector<glm::mat4>   inverseBindMatrices;
+    std::vector<Joint>       joints;
+    std::vector<std::string> jointNames;
+    std::vector<int>         parentIndex; // -1 = root, built once after loading
+};
+
+inline void buildParentIndex(Skeleton& skeleton) {
+    const std::size_t numJoints = skeleton.joints.size();
+    skeleton.parentIndex.assign(numJoints, -1);
+
+    // Build id -> slot map so we don't assume hierarchy[i] == joints[i]
+    std::unordered_map<JointId, std::size_t> idToSlot;
+    idToSlot.reserve(numJoints);
+    for (std::size_t j = 0; j < numJoints; ++j)
+        idToSlot[skeleton.joints[j].id] = j;
+
+    for (const auto& node : skeleton.hierarchy) {
+        auto parentIt = idToSlot.find(node.id);
+        if (parentIt == idToSlot.end()) continue;
+
+        for (const JointId childId : node.children) {
+            auto childIt = idToSlot.find(childId);
+            if (childIt != idToSlot.end())
+                skeleton.parentIndex[childIt->second] = static_cast<int>(parentIt->second);
+        }
+    }
+}
+
+#endif // SKELETON_H
--- a/destrum/include/destrum/Graphics/ids.h
+++ b/destrum/include/destrum/Graphics/ids.h
@@ -16,4 +16,8 @@ constexpr MaterialID NULL_MATERIAL_ID = std::numeric_limits<std::uint32_t>::max(
 using BindlessID = std::uint32_t;
 constexpr BindlessID NULL_BINDLESS_ID = std::numeric_limits<std::uint32_t>::max();

+using JointId = std::uint16_t;
+static const JointId NULL_JOINT_ID = std::numeric_limits<JointId>::max();
+static const JointId ROOT_JOINT_ID = 0;
+
 #endif //IDS_H
--- a/destrum/include/destrum/Util/MathUtils.h
+++ b/destrum/include/destrum/Util/MathUtils.h
@@ -1,4 +1,86 @@
 #ifndef MATHUTILS_H
 #define MATHUTILS_H

+#include <destrum/Graphics/Frustum.h>
+#include <span>
+#include <array>
+#include <glm/glm.hpp>
+#include <cassert>
+
+#include "glm/gtx/norm.hpp"
+
+inline Sphere calculateBoundingSphere(std::span<glm::vec3> positions)
+{
+    assert(!positions.empty());
+    auto calculateInitialSphere = [](const std::span<glm::vec3>& positions) -> Sphere {
+        constexpr int dirCount = 13;
+        static const std::array<glm::vec3, dirCount> direction = {{
+            {1.f, 0.f, 0.f},
+            {0.f, 1.f, 0.f},
+            {0.f, 0.f, 1.f},
+            {1.f, 1.f, 0.f},
+            {1.f, 0.f, 1.f},
+            {0.f, 1.f, 1.f},
+            {1.f, -1.f, 0.f},
+            {1.f, 0.f, -1.f},
+            {0.f, 1.f, -1.f},
+            {1.f, 1.f, 1.f},
+            {1.f, -1.f, 1.f},
+            {1.f, 1.f, -1.f},
+            {1.f, -1.f, -1.f},
+        }};
+
+        std::array<float, dirCount> dmin{};
+        std::array<float, dirCount> dmax{};
+        std::array<std::size_t, dirCount> imin{};
+        std::array<std::size_t, dirCount> imax{};
+
+        // Find min and max dot products for each direction and record vertex indices.
+        for (int j = 0; j < dirCount; ++j) {
+            const auto& u = direction[j];
+            dmin[j] = glm::dot(u, positions[0]);
+            dmax[j] = dmin[j];
+            for (std::size_t i = 1; i < positions.size(); ++i) {
+                const auto d = glm::dot(u, positions[i]);
+                if (d < dmin[j]) {
+                    dmin[j] = d;
+                    imin[j] = i;
+                } else if (d > dmax[j]) {
+                    dmax[j] = d;
+                    imax[j] = i;
+                }
+            };
+        }
+
+        // Find direction for which vertices at min and max extents are furthest apart.
+        float d2 = glm::length2(positions[imax[0]] - positions[imin[0]]);
+        int k = 0;
+        for (int j = 1; j < dirCount; j++) {
+            const auto m2 = glm::length2(positions[imax[j]] - positions[imin[j]]);
+            if (m2 > d2) {
+                d2 = m2;
+                k = j;
+            }
+        }
+
+        const auto center = (positions[imin[k]] + positions[imax[k]]) * 0.5f;
+        float radius = sqrt(d2) * 0.5f;
+        return {center, radius};
+    };
+
+    // Determine initial center and radius.
+    auto s = calculateInitialSphere(positions);
+    // Make pass through vertices and adjust sphere as necessary.
+    for (std::size_t i = 0; i < positions.size(); i++) {
+        const auto pv = positions[i] - s.center;
+        float m2 = glm::length2(pv);
+        if (m2 > s.radius * s.radius) {
+            auto q = s.center - (pv * (s.radius / std::sqrt(m2)));
+            s.center = (q + positions[i]) * 0.5f;
+            s.radius = glm::length(q - s.center);
+        }
+    }
+    return s;
+}
+
 #endif //MATHUTILS_H
--- a/destrum/include/destrum/Util/ModelLoader.h
+++ b/destrum/include/destrum/Util/ModelLoader.h
@@ -1,498 +1,521 @@
 #ifndef MODELLOADER_H
 #define MODELLOADER_H

-// CPUMesh loader with tinygltf
-// - Loads first scene (or default scene), iterates nodes, extracts mesh primitives.
-// - Handles POSITION/NORMAL/TANGENT/TEXCOORD_0 and indices.
-// - Computes minPos/maxPos.
-// - Can return per-primitive meshes, or merged-per-gltf-mesh meshes.
-// - IMPORTANT FIX: Applies node transforms (TRS / matrix) so models don't appear flipped/rotated.
+// CPUMesh loader using Assimp
+// - Loads first scene, iterates nodes recursively, extracts mesh primitives.
+// - Handles POSITION / NORMAL / TANGENT / TEXCOORD_0 and indices.
+// - Handles bones/skinning data and skeletal animations.
+// - Computes minPos / maxPos (in world space).
+// - Can return per-primitive meshes (PerPrimitive) or one merged mesh per
+//   Assimp mesh-node (MergedPerMesh).
+// - LoadSkinnedModel loads meshes + skeleton + animations in one call.
+// - For skinned meshes, vertex positions are kept in local/bind-pose space
+//   (identity world transform) so the skinning shader can apply joint matrices
+//   correctly at runtime.
 //
-// NOTE: Defining TINYGLTF_IMPLEMENTATION in a header can cause ODR / multiple-definition issues
-// if included in more than one translation unit. Best practice: put the TINYGLTF_IMPLEMENTATION
-// define in exactly one .cpp. I keep it here because your original file did, but you may want
-// to move it.
+// Link against:  assimp  (e.g. -lassimp)
+// Supports any format Assimp understands (.gltf, .glb, .fbx, .obj, …).

-#define TINYGLTF_IMPLEMENTATION
-#define TINYGLTF_NO_STB_IMAGE_WRITE
-#include <tiny_gltf.h>
+#include <assimp/Importer.hpp>
+#include <assimp/scene.h>
+#include <assimp/postprocess.h>

 #include <glm/glm.hpp>
-#include <glm/gtc/matrix_transform.hpp>  // translate/scale
-#include <glm/gtc/quaternion.hpp>        // quat
-#include <glm/gtx/quaternion.hpp>        // mat4_cast
+#include <glm/gtc/matrix_transform.hpp>
+#include <glm/gtc/type_ptr.hpp>
+#include <glm/gtc/quaternion.hpp>

 #include <destrum/Graphics/Resources/Mesh.h>
+#include <destrum/Graphics/Skeleton.h>
+#include <destrum/Graphics/SkeletalAnimation.h>

+#include <algorithm>
 #include <cstdint>
-#include <cstring>
 #include <limits>
-#include <optional>
+#include <stdexcept>
 #include <string>
 #include <unordered_map>
 #include <vector>
-#include <stdexcept>
-#include <iostream>
-#include <algorithm>

 namespace ModelLoader {

-    // -------------------- helpers --------------------
+// ─── Helpers ──────────────────────────────────────────────────────────────────

-    static size_t ComponentSizeInBytes(int componentType) {
-        switch (componentType) {
-            case TINYGLTF_COMPONENT_TYPE_BYTE: return 1;
-            case TINYGLTF_COMPONENT_TYPE_UNSIGNED_BYTE: return 1;
-            case TINYGLTF_COMPONENT_TYPE_SHORT: return 2;
-            case TINYGLTF_COMPONENT_TYPE_UNSIGNED_SHORT: return 2;
-            case TINYGLTF_COMPONENT_TYPE_INT: return 4;
-            case TINYGLTF_COMPONENT_TYPE_UNSIGNED_INT: return 4;
-            case TINYGLTF_COMPONENT_TYPE_FLOAT: return 4;
-            case TINYGLTF_COMPONENT_TYPE_DOUBLE: return 8;
-            default: throw std::runtime_error("Unknown glTF component type");
-        }
-    }
+// Convert an Assimp row-major 4×4 matrix to GLM column-major.
+static glm::mat4 ToGLM(const aiMatrix4x4& m) {
+    return glm::transpose(glm::make_mat4(&m.a1));
+}

-    static int TypeNumComponents(int type) {
-        switch (type) {
-            case TINYGLTF_TYPE_SCALAR: return 1;
-            case TINYGLTF_TYPE_VEC2: return 2;
-            case TINYGLTF_TYPE_VEC3: return 3;
-            case TINYGLTF_TYPE_VEC4: return 4;
-            case TINYGLTF_TYPE_MAT2: return 4;
-            case TINYGLTF_TYPE_MAT3: return 9;
-            case TINYGLTF_TYPE_MAT4: return 16;
-            default: throw std::runtime_error("Unknown glTF type");
-        }
-    }
+static void UpdateBounds(glm::vec3& mn, glm::vec3& mx, const glm::vec3& p) {
+    mn.x = std::min(mn.x, p.x);
+    mn.y = std::min(mn.y, p.y);
+    mn.z = std::min(mn.z, p.z);
+    mx.x = std::max(mx.x, p.x);
+    mx.y = std::max(mx.y, p.y);
+    mx.z = std::max(mx.z, p.z);
+}

-    // Returns pointer to the first element, plus stride in bytes.
-    static std::pair<const std::uint8_t*, size_t>
-    GetAccessorDataPtrAndStride(const tinygltf::Model& model, const tinygltf::Accessor& accessor) {
-        if (accessor.bufferView < 0) {
-            throw std::runtime_error("Accessor has no bufferView");
-        }
-        const tinygltf::BufferView& bv = model.bufferViews.at(accessor.bufferView);
-        const tinygltf::Buffer& buf = model.buffers.at(bv.buffer);
+// ─── Post-process flags ───────────────────────────────────────────────────────

-        const size_t componentSize = ComponentSizeInBytes(accessor.componentType);
-        const int numComps = TypeNumComponents(accessor.type);
-        const size_t packedStride = componentSize * size_t(numComps);
+static constexpr unsigned int kImportFlags =
+    aiProcess_Triangulate           |
+    aiProcess_CalcTangentSpace      |
+    aiProcess_JoinIdenticalVertices |
+    aiProcess_GenSmoothNormals      |
+    aiProcess_FlipUVs               |
+    aiProcess_LimitBoneWeights;     // guarantees <= 4 weights per vertex

-        size_t stride = bv.byteStride ? size_t(bv.byteStride) : packedStride;
+static const aiScene* LoadScene(Assimp::Importer& importer, const std::string& path) {
+    const aiScene* scene = importer.ReadFile(path, kImportFlags);
+    if (!scene || (scene->mFlags & AI_SCENE_FLAGS_INCOMPLETE) || !scene->mRootNode)
+        throw std::runtime_error(std::string("Assimp failed to load: ") + importer.GetErrorString());
+    return scene;
+}

-        const size_t start = size_t(bv.byteOffset) + size_t(accessor.byteOffset);
-        // Conservative bounds check; don't hard-fail on weird stride but catches obvious issues.
-        if (start > buf.data.size()) {
-            throw std::runtime_error("Accessor start is out of buffer bounds");
-        }
+// ─── Primitive extraction ─────────────────────────────────────────────────────

-        const std::uint8_t* ptr = buf.data.data() + start;
-        return {ptr, stride};
-    }
+// Build one CPUMesh from a single aiMesh, applying the given world transform.
+// For skinned meshes, pass glm::mat4{1.f} so vertices stay in bind-pose space.
+// Skinning data is NOT populated here — call LoadSkinningData afterwards.
+static CPUMesh LoadAiMeshIntoCPUMesh(const aiMesh* mesh,
+                                      const std::string& name,
+                                      const glm::mat4& world) {
+    CPUMesh out{};
+    out.name = name;

-    template <typename T>
-    static T ReadAs(const std::uint8_t* p) {
-        T v{};
-        std::memcpy(&v, p, sizeof(T));
-        return v;
-    }
+    const size_t vertexCount = mesh->mNumVertices;
+    out.vertices.resize(vertexCount);

-    static glm::vec3 ReadVec3Float(const std::uint8_t* base, size_t stride, size_t i) {
-        const std::uint8_t* p = base + i * stride;
-        const float x = ReadAs<float>(p + 0);
-        const float y = ReadAs<float>(p + 4);
-        const float z = ReadAs<float>(p + 8);
-        return glm::vec3{x, y, z};
-    }
+    const glm::mat3 nrmMat = glm::transpose(glm::inverse(glm::mat3(world)));
+    const glm::mat3 tanMat = glm::mat3(world);

-    static glm::vec2 ReadVec2Float(const std::uint8_t* base, size_t stride, size_t i) {
-        const std::uint8_t* p = base + i * stride;
-        const float x = ReadAs<float>(p + 0);
-        const float y = ReadAs<float>(p + 4);
-        return glm::vec2{x, y};
-    }
+    glm::vec3 mn{ std::numeric_limits<float>::infinity()};
+    glm::vec3 mx{-std::numeric_limits<float>::infinity()};

-    static glm::vec4 ReadVec4Float(const std::uint8_t* base, size_t stride, size_t i) {
-        const std::uint8_t* p = base + i * stride;
-        const float x = ReadAs<float>(p + 0);
-        const float y = ReadAs<float>(p + 4);
-        const float z = ReadAs<float>(p + 8);
-        const float w = ReadAs<float>(p + 12);
-        return glm::vec4{x, y, z, w};
-    }
+    for (size_t i = 0; i < vertexCount; ++i) {
+        CPUMesh::Vertex v{};

-    static std::uint32_t ReadIndexAsU32(const tinygltf::Model& model, const tinygltf::Accessor& accessor, size_t i) {
-        auto [base, stride] = GetAccessorDataPtrAndStride(model, accessor);
-        const std::uint8_t* p = base + i * stride;
+        // Position (required)
+        const aiVector3D& ap = mesh->mVertices[i];
+        v.position = glm::vec3(world * glm::vec4(ap.x, ap.y, ap.z, 1.0f));
+        UpdateBounds(mn, mx, v.position);

-        switch (accessor.componentType) {
-            case TINYGLTF_COMPONENT_TYPE_UNSIGNED_BYTE:
-                return static_cast<std::uint32_t>(ReadAs<std::uint8_t>(p));
-            case TINYGLTF_COMPONENT_TYPE_UNSIGNED_SHORT:
-                return static_cast<std::uint32_t>(ReadAs<std::uint16_t>(p));
-            case TINYGLTF_COMPONENT_TYPE_UNSIGNED_INT:
-                return static_cast<std::uint32_t>(ReadAs<std::uint32_t>(p));
-            default:
-                throw std::runtime_error("Unsupported index componentType (expected u8/u16/u32)");
-        }
-    }
-
-    static void UpdateBounds(glm::vec3& mn, glm::vec3& mx, const glm::vec3& p) {
-        mn.x = std::min(mn.x, p.x);
-        mn.y = std::min(mn.y, p.y);
-        mn.z = std::min(mn.z, p.z);
-        mx.x = std::max(mx.x, p.x);
-        mx.y = std::max(mx.y, p.y);
-        mx.z = std::max(mx.z, p.z);
-    }
-
-    // Build a node local matrix from either node.matrix or TRS.
-    // glTF stores rotation as quaternion [x,y,z,w].
-    static glm::mat4 NodeLocalMatrix(const tinygltf::Node& node) {
-        if (node.matrix.size() == 16) {
-            glm::mat4 m(1.0f);
-            // glTF is column-major; GLM is column-major -> fill columns.
-            for (int c = 0; c < 4; ++c)
-                for (int r = 0; r < 4; ++r)
-                    m[c][r] = static_cast<float>(node.matrix[c * 4 + r]);
-            return m;
-        }
-
-        glm::vec3 t(0.0f);
-        if (node.translation.size() == 3) {
-            t = glm::vec3(
-                static_cast<float>(node.translation[0]),
-                static_cast<float>(node.translation[1]),
-                static_cast<float>(node.translation[2])
-            );
-        }
-
-        glm::quat q(1.0f, 0.0f, 0.0f, 0.0f); // w,x,y,z
-        if (node.rotation.size() == 4) {
-            q = glm::quat(
-                static_cast<float>(node.rotation[3]), // w
-                static_cast<float>(node.rotation[0]), // x
-                static_cast<float>(node.rotation[1]), // y
-                static_cast<float>(node.rotation[2])  // z
-            );
-        }
-
-        glm::vec3 s(1.0f);
-        if (node.scale.size() == 3) {
-            s = glm::vec3(
-                static_cast<float>(node.scale[0]),
-                static_cast<float>(node.scale[1]),
-                static_cast<float>(node.scale[2])
-            );
-        }
-
-        const glm::mat4 T = glm::translate(glm::mat4(1.0f), t);
-        const glm::mat4 R = glm::toMat4(q);
-        const glm::mat4 S = glm::scale(glm::mat4(1.0f), s);
-        return T * R * S;
-    }
-
-    // -------------------- primitive extraction --------------------
-
-    static CPUMesh LoadPrimitiveIntoCPUMesh(const tinygltf::Model& model,
-                                           const tinygltf::Primitive& prim,
-                                           const std::string& nameForMesh,
-                                           const glm::mat4& world) {
-        CPUMesh out{};
-        out.name = nameForMesh;
-
-        // POSITION is required
-        auto itPos = prim.attributes.find("POSITION");
-        if (itPos == prim.attributes.end()) {
-            throw std::runtime_error("Primitive has no POSITION attribute");
-        }
-
-        const tinygltf::Accessor& accPos = model.accessors.at(itPos->second);
-        if (accPos.componentType != TINYGLTF_COMPONENT_TYPE_FLOAT || accPos.type != TINYGLTF_TYPE_VEC3) {
-            throw std::runtime_error("POSITION must be VEC3 float for this loader");
-        }
-
-        const size_t vertexCount = size_t(accPos.count);
-
-        // Optional attributes
-        const tinygltf::Accessor* accNormal = nullptr;
-        const tinygltf::Accessor* accTangent = nullptr;
-        const tinygltf::Accessor* accUV0 = nullptr;
-
-        if (auto it = prim.attributes.find("NORMAL"); it != prim.attributes.end()) {
-            accNormal = &model.accessors.at(it->second);
-            if (accNormal->componentType != TINYGLTF_COMPONENT_TYPE_FLOAT || accNormal->type != TINYGLTF_TYPE_VEC3)
-                accNormal = nullptr;
-        }
-        if (auto it = prim.attributes.find("TANGENT"); it != prim.attributes.end()) {
-            accTangent = &model.accessors.at(it->second);
-            if (accTangent->componentType != TINYGLTF_COMPONENT_TYPE_FLOAT || accTangent->type != TINYGLTF_TYPE_VEC4)
-                accTangent = nullptr;
-        }
-        if (auto it = prim.attributes.find("TEXCOORD_0"); it != prim.attributes.end()) {
-            accUV0 = &model.accessors.at(it->second);
-            if (accUV0->componentType != TINYGLTF_COMPONENT_TYPE_FLOAT || accUV0->type != TINYGLTF_TYPE_VEC2)
-                accUV0 = nullptr;
-        }
-
-        // Prepare pointers/strides
-        auto [posBase, posStride] = GetAccessorDataPtrAndStride(model, accPos);
-
-        const std::uint8_t* nrmBase = nullptr;
-        size_t nrmStride = 0;
-        const std::uint8_t* tanBase = nullptr;
-        size_t tanStride = 0;
-        const std::uint8_t* uvBase = nullptr;
-        size_t uvStride = 0;
-
-        if (accNormal) {
-            auto p = GetAccessorDataPtrAndStride(model, *accNormal);
-            nrmBase = p.first;
-            nrmStride = p.second;
-            if (size_t(accNormal->count) != vertexCount) accNormal = nullptr;
-        }
-        if (accTangent) {
-            auto p = GetAccessorDataPtrAndStride(model, *accTangent);
-            tanBase = p.first;
-            tanStride = p.second;
-            if (size_t(accTangent->count) != vertexCount) accTangent = nullptr;
-        }
-        if (accUV0) {
-            auto p = GetAccessorDataPtrAndStride(model, *accUV0);
-            uvBase = p.first;
-            uvStride = p.second;
-            if (size_t(accUV0->count) != vertexCount) accUV0 = nullptr;
-        }
-
-        // Allocate vertices
-        out.vertices.resize(vertexCount);
-
-        // Bounds init (in world space, because we transform positions)
-        glm::vec3 mn{std::numeric_limits<float>::infinity()};
-        glm::vec3 mx{-std::numeric_limits<float>::infinity()};
-
-        // Normal matrix
-        const glm::mat3 nrmMat = glm::transpose(glm::inverse(glm::mat3(world)));
-        const glm::mat3 tanMat = glm::mat3(world);
-
-        for (size_t i = 0; i < vertexCount; ++i) {
-            CPUMesh::Vertex v{};
-
-            // Position
-            const glm::vec3 pLocal = ReadVec3Float(posBase, posStride, i);
-            v.position = glm::vec3(world * glm::vec4(pLocal, 1.0f));
-            UpdateBounds(mn, mx, v.position);
-
-            // Normal
-            if (accNormal) {
-                const glm::vec3 nLocal = ReadVec3Float(nrmBase, nrmStride, i);
-                v.normal = glm::normalize(nrmMat * nLocal);
-            } else {
-                v.normal = glm::vec3(0.0f, 1.0f, 0.0f);
-            }
-
-            // UV
-            if (accUV0) {
-                glm::vec2 uv = ReadVec2Float(uvBase, uvStride, i);
-                v.uv_x = uv.x;
-                v.uv_y = uv.y;
-            } else {
-                v.uv_x = 0.0f;
-                v.uv_y = 0.0f;
-            }
-
-            // Tangent
-            if (accTangent) {
-                glm::vec4 t = ReadVec4Float(tanBase, tanStride, i);
-                glm::vec3 t3 = glm::normalize(tanMat * glm::vec3(t));
-                v.tangent = glm::vec4(t3, t.w); // keep handedness in w
-            } else {
-                v.tangent = glm::vec4(1.0f, 0.0f, 0.0f, 1.0f);
-            }
-
-            out.vertices[i] = v;
-        }
-
-        out.minPos = (vertexCount > 0) ? mn : glm::vec3(0.0f);
-        out.maxPos = (vertexCount > 0) ? mx : glm::vec3(0.0f);
-
-        // Indices
-        if (prim.indices >= 0) {
-            const tinygltf::Accessor& accIdx = model.accessors.at(prim.indices);
-            if (accIdx.type != TINYGLTF_TYPE_SCALAR) {
-                throw std::runtime_error("Indices accessor must be SCALAR");
-            }
-
-            out.indices.resize(size_t(accIdx.count));
-            for (size_t i = 0; i < size_t(accIdx.count); ++i) {
-                out.indices[i] = ReadIndexAsU32(model, accIdx, i);
-            }
+        // Normal
+        if (mesh->HasNormals()) {
+            const aiVector3D& an = mesh->mNormals[i];
+            v.normal = glm::normalize(nrmMat * glm::vec3(an.x, an.y, an.z));
        } else {
-            out.indices.resize(vertexCount);
-            for (size_t i = 0; i < vertexCount; ++i) out.indices[i] = static_cast<std::uint32_t>(i);
+            v.normal = glm::vec3(0.0f, 1.0f, 0.0f);
        }

-        // If your renderer expects opposite winding vs glTF, you can flip triangle order.
-        // Keep what you had:
-        // for (size_t i = 0; i + 2 < out.indices.size(); i += 3) {
-        //     std::swap(out.indices[i + 1], out.indices[i + 2]);
-        // }
-
-        return out;
-    }
-
-    // Merge "src" into "dst" (offset indices)
-    static void AppendMesh(CPUMesh& dst, const CPUMesh& src) {
-        const std::uint32_t baseVertex = static_cast<std::uint32_t>(dst.vertices.size());
-        dst.vertices.insert(dst.vertices.end(), src.vertices.begin(), src.vertices.end());
-
-        dst.indices.reserve(dst.indices.size() + src.indices.size());
-        for (std::uint32_t idx : src.indices) {
-            dst.indices.push_back(baseVertex + idx);
-        }
-
-        if (dst.vertices.size() == src.vertices.size()) {
-            dst.minPos = src.minPos;
-            dst.maxPos = src.maxPos;
+        // UV channel 0
+        if (mesh->HasTextureCoords(0)) {
+            v.uv_x = mesh->mTextureCoords[0][i].x;
+            v.uv_y = mesh->mTextureCoords[0][i].y;
        } else {
-            dst.minPos = glm::vec3(
-                std::min(dst.minPos.x, src.minPos.x),
-                std::min(dst.minPos.y, src.minPos.y),
-                std::min(dst.minPos.z, src.minPos.z)
-            );
-            dst.maxPos = glm::vec3(
-                std::max(dst.maxPos.x, src.maxPos.x),
-                std::max(dst.maxPos.y, src.maxPos.y),
-                std::max(dst.maxPos.z, src.maxPos.z)
-            );
+            v.uv_x = 0.0f;
+            v.uv_y = 0.0f;
+        }
+
+        // Tangent + bitangent sign (stored in w as +1/-1 handedness)
+        if (mesh->HasTangentsAndBitangents()) {
+            const aiVector3D& at = mesh->mTangents[i];
+            const aiVector3D& ab = mesh->mBitangents[i];
+            glm::vec3 t3 = glm::normalize(tanMat * glm::vec3(at.x, at.y, at.z));
+            glm::vec3 b3 = glm::normalize(tanMat * glm::vec3(ab.x, ab.y, ab.z));
+            glm::vec3 n3 = v.normal;
+            float sign = (glm::dot(glm::cross(n3, t3), b3) < 0.0f) ? -1.0f : 1.0f;
+            v.tangent = glm::vec4(t3, sign);
+        } else {
+            v.tangent = glm::vec4(1.0f, 0.0f, 0.0f, 1.0f);
+        }
+
+        out.vertices[i] = v;
+    }
+
+    out.minPos = (vertexCount > 0) ? mn : glm::vec3(0.0f);
+    out.maxPos = (vertexCount > 0) ? mx : glm::vec3(0.0f);
+
+    // Indices
+    if (mesh->HasFaces()) {
+        out.indices.reserve(size_t(mesh->mNumFaces) * 3);
+        for (unsigned int f = 0; f < mesh->mNumFaces; ++f) {
+            const aiFace& face = mesh->mFaces[f];
+            for (unsigned int k = 0; k < face.mNumIndices; ++k)
+                out.indices.push_back(static_cast<std::uint32_t>(face.mIndices[k]));
+        }
+    } else {
+        out.indices.resize(vertexCount);
+        for (size_t i = 0; i < vertexCount; ++i)
+            out.indices[i] = static_cast<std::uint32_t>(i);
+    }
+
+    return out;
+}
+
+// Merge src into dst, offsetting src indices by the current dst vertex count.
+static void AppendMesh(CPUMesh& dst, const CPUMesh& src) {
+    const std::uint32_t base = static_cast<std::uint32_t>(dst.vertices.size());
+    dst.vertices.insert(dst.vertices.end(), src.vertices.begin(), src.vertices.end());
+
+    dst.indices.reserve(dst.indices.size() + src.indices.size());
+    for (std::uint32_t idx : src.indices)
+        dst.indices.push_back(base + idx);
+
+    dst.minPos = glm::vec3(
+        std::min(dst.minPos.x, src.minPos.x),
+        std::min(dst.minPos.y, src.minPos.y),
+        std::min(dst.minPos.z, src.minPos.z)
+    );
+    dst.maxPos = glm::vec3(
+        std::max(dst.maxPos.x, src.maxPos.x),
+        std::max(dst.maxPos.y, src.maxPos.y),
+        std::max(dst.maxPos.z, src.maxPos.z)
+    );
+}
+
+// ─── Skeleton loading ─────────────────────────────────────────────────────────
+
+static Skeleton LoadSkeleton(const aiScene* scene) {
+    Skeleton skeleton;
+
+    // Pass 1: collect all bone names and their inverse bind matrices
+    // from every mesh in the scene.
+    std::unordered_map<std::string, glm::mat4> boneOffsets;
+    for (unsigned int m = 0; m < scene->mNumMeshes; ++m) {
+        const aiMesh* mesh = scene->mMeshes[m];
+        for (unsigned int b = 0; b < mesh->mNumBones; ++b) {
+            const aiBone* bone = mesh->mBones[b];
+            std::string boneName(bone->mName.C_Str());
+            if (!boneOffsets.count(boneName))
+                boneOffsets[boneName] = ToGLM(bone->mOffsetMatrix);
        }
    }

-    // -------------------- public API --------------------
+    if (boneOffsets.empty()) return skeleton; // static mesh, no skeleton

-    struct NodeStackItem {
-        int nodeIdx;
-        glm::mat4 parentWorld;
-    };
+    // Pass 2: walk the node tree in depth-first order (parent before child),
+    // registering only nodes that correspond to actual bones.
+    // This guarantees joints are ordered parent-before-child, which is required
+    // by Animator::computeJointMatrices.
+    JointId nextId = 0;

-    // Option A: return one CPUMesh per *primitive* encountered in the scene
-    static std::vector<CPUMesh> LoadGLTF_CPUMeshes_PerPrimitive(const std::string& path) {
-        tinygltf::TinyGLTF loader;
-        tinygltf::Model model;
-        std::string err, warn;
+    struct StackItem { const aiNode* node; int parentIdx; };
+    std::vector<StackItem> stack{{ scene->mRootNode, -1 }};

-        bool ok = false;
-        const bool isGLB = (path.size() >= 4 && path.substr(path.size() - 4) == ".glb");
-        if (isGLB) ok = loader.LoadBinaryFromFile(&model, &err, &warn, path);
-        else ok = loader.LoadASCIIFromFile(&model, &err, &warn, path);
+    while (!stack.empty()) {
+        auto [node, parentIdx] = stack.back();
+        stack.pop_back();

-        if (!warn.empty()) std::cerr << "tinygltf warn: " << warn << "\n";
-        if (!ok) throw std::runtime_error("Failed to load glTF: " + err);
+        std::string name(node->mName.C_Str());
+        int myIdx = parentIdx; // default: pass parent through to children

-        int sceneIndex = model.defaultScene >= 0 ? model.defaultScene : 0;
-        if (sceneIndex < 0 || sceneIndex >= int(model.scenes.size())) {
-            throw std::runtime_error("glTF has no valid scene");
+        if (boneOffsets.count(name)) {
+            const JointId id = nextId++;
+            myIdx = static_cast<int>(id);
+
+            Joint joint{};
+            joint.id = id;
+            skeleton.joints.push_back(joint);
+            skeleton.jointNames.push_back(name);
+            skeleton.inverseBindMatrices.push_back(boneOffsets[name]);
+
+            // Grow hierarchy to accommodate this id
+            while (skeleton.hierarchy.size() <= id)
+                skeleton.hierarchy.push_back({});
+
+            skeleton.hierarchy[id].id = id; // record the node's own id
+
+            if (parentIdx >= 0)
+                skeleton.hierarchy[parentIdx].children.push_back(id);
        }

-        std::vector<CPUMesh> result;
-        const tinygltf::Scene& scene = model.scenes.at(sceneIndex);
+        // Push children in reverse so left-most child is processed first
+        for (int c = static_cast<int>(node->mNumChildren) - 1; c >= 0; --c)
+            stack.push_back({ node->mChildren[c], myIdx });
+    }

-        std::vector<NodeStackItem> stack;
-        stack.reserve(scene.nodes.size());
-        for (int n : scene.nodes) stack.push_back({n, glm::mat4(1.0f)});
+    buildParentIndex(skeleton);
+    return skeleton;
+}

-        while (!stack.empty()) {
-            NodeStackItem it = stack.back();
-            stack.pop_back();
+// ─── Skinning data ────────────────────────────────────────────────────────────

-            const tinygltf::Node& node = model.nodes.at(it.nodeIdx);
-            const glm::mat4 local = NodeLocalMatrix(node);
-            const glm::mat4 world = it.parentWorld * local;
+// Populate cpuMesh.skinningData from the matching aiMesh.
+// Must be called after LoadAiMeshIntoCPUMesh so vertices are already sized.
+static void LoadSkinningData(CPUMesh& cpuMesh,
+                              const aiMesh* aiMesh,
+                              const Skeleton& skeleton) {
+    if (!aiMesh->HasBones() || skeleton.joints.empty()) return;

-            for (int child : node.children) stack.push_back({child, world});
+    cpuMesh.skinningData.assign(cpuMesh.vertices.size(),
+        CPUMesh::SkinningData{ {0, 0, 0, 0}, {0.f, 0.f, 0.f, 0.f} });

-            if (node.mesh < 0) continue;
-            const tinygltf::Mesh& mesh = model.meshes.at(node.mesh);
+    std::vector<int> weightCount(cpuMesh.vertices.size(), 0);

-            for (size_t p = 0; p < mesh.primitives.size(); ++p) {
-                const tinygltf::Primitive& prim = mesh.primitives[p];
-                CPUMesh cpu = LoadPrimitiveIntoCPUMesh(
-                    model,
-                    prim,
-                    mesh.name.empty() ? ("mesh_" + std::to_string(node.mesh)) : mesh.name,
-                    world
-                );
-                cpu.name += "_prim" + std::to_string(p);
-                result.push_back(std::move(cpu));
+    for (unsigned int b = 0; b < aiMesh->mNumBones; ++b) {
+        const aiBone* bone = aiMesh->mBones[b];
+        std::string boneName(bone->mName.C_Str());
+
+        // Find which joint index this bone maps to
+        std::uint32_t jointIdx = 0;
+        bool found = false;
+        for (std::size_t j = 0; j < skeleton.jointNames.size(); ++j) {
+            if (skeleton.jointNames[j] == boneName) {
+                jointIdx = static_cast<std::uint32_t>(j);
+                found = true;
+                break;
            }
        }
+        if (!found) continue;

-        return result;
-    }
+        for (unsigned int w = 0; w < bone->mNumWeights; ++w) {
+            const unsigned int vertIdx = bone->mWeights[w].mVertexId;
+            const float        weight  = bone->mWeights[w].mWeight;
+            const int          slot    = weightCount[vertIdx];

-    // Option B: return one CPUMesh per *glTF mesh instance*, merging all primitives of that mesh into one CPUMesh.
-    // Note: If the same glTF mesh is instanced by multiple nodes, you'll get one merged CPUMesh PER NODE instance.
-    static std::vector<CPUMesh> LoadGLTF_CPUMeshes_MergedPerMesh(const std::string& path) {
-        tinygltf::TinyGLTF loader;
-        tinygltf::Model model;
-        std::string err, warn;
+            if (slot >= 4) continue; // aiProcess_LimitBoneWeights should prevent this

-        bool ok = false;
-        const bool isGLB = (path.size() >= 4 && path.substr(path.size() - 4) == ".glb");
-        if (isGLB) ok = loader.LoadBinaryFromFile(&model, &err, &warn, path);
-        else ok = loader.LoadASCIIFromFile(&model, &err, &warn, path);
-
-        if (!warn.empty()) std::cerr << "tinygltf warn: " << warn << "\n";
-        if (!ok) throw std::runtime_error("Failed to load glTF: " + err);
-
-        int sceneIndex = model.defaultScene >= 0 ? model.defaultScene : 0;
-        if (sceneIndex < 0 || sceneIndex >= int(model.scenes.size())) {
-            throw std::runtime_error("glTF has no valid scene");
+            cpuMesh.skinningData[vertIdx].jointIds[slot] = jointIdx;
+            cpuMesh.skinningData[vertIdx].weights[slot]  = weight;
+            ++weightCount[vertIdx];
        }
+    }
+}

-        std::vector<CPUMesh> result;
+// ─── Animation loading ────────────────────────────────────────────────────────

-        const tinygltf::Scene& scene = model.scenes.at(sceneIndex);
-        std::vector<NodeStackItem> stack;
-        stack.reserve(scene.nodes.size());
-        for (int n : scene.nodes) stack.push_back({n, glm::mat4(1.0f)});
+static std::vector<SkeletalAnimation> LoadAnimations(const aiScene* scene,
+                                                      const Skeleton& skeleton) {
+    std::vector<SkeletalAnimation> result;
+    if (!scene->HasAnimations()) return result;

-        while (!stack.empty()) {
-            NodeStackItem it = stack.back();
-            stack.pop_back();
+    for (unsigned int a = 0; a < scene->mNumAnimations; ++a) {
+        const aiAnimation* aiAnim = scene->mAnimations[a];
+        const double tps = aiAnim->mTicksPerSecond > 0.0 ? aiAnim->mTicksPerSecond : 25.0;

-            const tinygltf::Node& node = model.nodes.at(it.nodeIdx);
-            const glm::mat4 local = NodeLocalMatrix(node);
-            const glm::mat4 world = it.parentWorld * local;
+        SkeletalAnimation anim;
+        anim.name     = aiAnim->mName.C_Str();
+        anim.duration = static_cast<float>(aiAnim->mDuration / tps);
+        anim.looped   = true;

-            for (int child : node.children) stack.push_back({child, world});
+        for (unsigned int c = 0; c < aiAnim->mNumChannels; ++c) {
+            const aiNodeAnim* ch = aiAnim->mChannels[c];
+            std::string boneName(ch->mNodeName.C_Str());

-            if (node.mesh < 0) continue;
-            const tinygltf::Mesh& mesh = model.meshes.at(node.mesh);
-
-            CPUMesh merged{};
-            merged.name = mesh.name.empty() ? ("mesh_" + std::to_string(node.mesh)) : mesh.name;
-            merged.minPos = glm::vec3(std::numeric_limits<float>::infinity());
-            merged.maxPos = glm::vec3(-std::numeric_limits<float>::infinity());
-
-            bool any = false;
-            for (const tinygltf::Primitive& prim : mesh.primitives) {
-                CPUMesh part = LoadPrimitiveIntoCPUMesh(model, prim, merged.name, world);
-                if (!any) {
-                    merged = std::move(part);
-                    any = true;
-                } else {
-                    AppendMesh(merged, part);
+            // Map bone name to joint index
+            std::uint32_t jointIdx = 0;
+            bool found = false;
+            for (std::size_t j = 0; j < skeleton.jointNames.size(); ++j) {
+                if (skeleton.jointNames[j] == boneName) {
+                    jointIdx = static_cast<std::uint32_t>(j);
+                    found = true;
+                    break;
                }
            }
+            if (!found) continue;

-            if (any) result.push_back(std::move(merged));
+            SkeletalAnimation::Track track;
+            track.jointIndex = jointIdx;
+
+            // FIX: position, rotation and scale channels can have different
+            // keyframe counts and independent time axes. Build the track from
+            // the position channel's timeline and pick the nearest rotation/
+            // scale key for each sample rather than assuming index alignment.
+            // This avoids corrupted keyframes when counts differ.
+            const unsigned int numPos = ch->mNumPositionKeys;
+            const unsigned int numRot = ch->mNumRotationKeys;
+            const unsigned int numSca = ch->mNumScalingKeys;
+
+            // Use position keys to drive the timeline (most common case).
+            // If there are no position keys fall back to rotation keys.
+            const unsigned int numKeys = numPos > 0 ? numPos : numRot;
+
+            for (unsigned int k = 0; k < numKeys; ++k) {
+                SkeletalAnimation::Keyframe kf;
+
+                // Time comes from whichever channel drives this loop
+                if (numPos > 0) {
+                    kf.time = static_cast<float>(ch->mPositionKeys[k].mTime / tps);
+                    const auto& p = ch->mPositionKeys[k].mValue;
+                    kf.translation = { p.x, p.y, p.z };
+                } else {
+                    kf.time        = static_cast<float>(ch->mRotationKeys[k].mTime / tps);
+                    kf.translation = { 0.f, 0.f, 0.f };
+                }
+
+                // Nearest rotation key at this index
+                {
+                    const unsigned int ri = std::min(k, numRot - 1);
+                    const auto& r = ch->mRotationKeys[ri].mValue;
+                    kf.rotation = glm::quat(r.w, r.x, r.y, r.z); // glm: (w,x,y,z)
+                }
+
+                // Nearest scale key at this index
+                {
+                    const unsigned int si = std::min(k, numSca - 1);
+                    const auto& s = ch->mScalingKeys[si].mValue;
+                    kf.scale = { s.x, s.y, s.z };
+                }
+
+                track.keyframes.push_back(kf);
+            }
+
+            anim.tracks.push_back(std::move(track));
        }

-        return result;
+        result.push_back(std::move(anim));
    }

+    return result;
+}
+
+// ─── Public API: static meshes ────────────────────────────────────────────────
+
+// Option A: one CPUMesh per aiMesh reference encountered in the node tree.
+static std::vector<CPUMesh> LoadGLTF_CPUMeshes_PerPrimitive(const std::string& path) {
+    Assimp::Importer importer;
+    const aiScene* scene = LoadScene(importer, path);
+
+    std::vector<CPUMesh> result;
+
+    struct StackItem { const aiNode* node; glm::mat4 parentWorld; };
+    std::vector<StackItem> stack{{ scene->mRootNode, glm::mat4(1.0f) }};
+
+    while (!stack.empty()) {
+        auto [node, parentWorld] = stack.back();
+        stack.pop_back();
+
+        const glm::mat4 world = parentWorld * ToGLM(node->mTransformation);
+
+        for (unsigned int c = 0; c < node->mNumChildren; ++c)
+            stack.push_back({ node->mChildren[c], world });
+
+        for (unsigned int m = 0; m < node->mNumMeshes; ++m) {
+            const aiMesh* mesh = scene->mMeshes[node->mMeshes[m]];
+            std::string name = mesh->mName.length > 0
+                ? std::string(mesh->mName.C_Str())
+                : ("mesh_" + std::to_string(node->mMeshes[m]));
+            name += "_prim" + std::to_string(m);
+
+            result.push_back(LoadAiMeshIntoCPUMesh(mesh, name, world));
+        }
+    }
+
+    return result;
+}
+
+// Option B: one CPUMesh per node, merging all aiMesh references of that node.
+static std::vector<CPUMesh> LoadGLTF_CPUMeshes_MergedPerMesh(const std::string& path) {
+    Assimp::Importer importer;
+    const aiScene* scene = LoadScene(importer, path);
+
+    std::vector<CPUMesh> result;
+
+    struct StackItem { const aiNode* node; glm::mat4 parentWorld; };
+    std::vector<StackItem> stack{{ scene->mRootNode, glm::mat4(1.0f) }};
+
+    while (!stack.empty()) {
+        auto [node, parentWorld] = stack.back();
+        stack.pop_back();
+
+        const glm::mat4 world = parentWorld * ToGLM(node->mTransformation);
+
+        for (unsigned int c = 0; c < node->mNumChildren; ++c)
+            stack.push_back({ node->mChildren[c], world });
+
+        if (node->mNumMeshes == 0) continue;
+
+        std::string mergedName = node->mName.length > 0
+            ? std::string(node->mName.C_Str())
+            : std::string(scene->mMeshes[node->mMeshes[0]]->mName.C_Str());
+
+        CPUMesh merged{};
+        merged.name   = mergedName;
+        merged.minPos = glm::vec3( std::numeric_limits<float>::infinity());
+        merged.maxPos = glm::vec3(-std::numeric_limits<float>::infinity());
+
+        bool any = false;
+        for (unsigned int m = 0; m < node->mNumMeshes; ++m) {
+            const aiMesh* mesh = scene->mMeshes[node->mMeshes[m]];
+            CPUMesh part = LoadAiMeshIntoCPUMesh(mesh, mergedName, world);
+            if (!any) { merged = std::move(part); any = true; }
+            else       AppendMesh(merged, part);
+        }
+
+        if (any) result.push_back(std::move(merged));
+    }
+
+    return result;
+}
+
+// ─── Public API: skinned model ────────────────────────────────────────────────
+
+struct SkinnedModel {
+    std::vector<CPUMesh>           meshes;
+    Skeleton                       skeleton;
+    std::vector<SkeletalAnimation> animations;
+};
+
+// Loads meshes + skeleton + animations in a single Assimp pass.
+// Each CPUMesh has skinningData and skeleton populated.
+//
+// IMPORTANT: skinned mesh vertices are loaded with an identity world transform
+// so they remain in bind-pose space. The skinning shader applies joint matrices
+// at runtime — baking the node transform into positions would break that.
+//
+// Usage:
+//   auto model = ModelLoader::LoadSkinnedModel("character.glb");
+//   MeshId id  = meshCache.uploadMesh(model.meshes[0]);
+//   auto* anim = go.AddComponent<Animator>();
+//   for (auto& clip : model.animations)
+//       anim->addClip(std::make_shared<SkeletalAnimation>(std::move(clip)));
+//   anim->play("Walk");
+static SkinnedModel LoadSkinnedModel(const std::string& path) {
+    Assimp::Importer importer;
+    const aiScene* scene = LoadScene(importer, path);
+
+    SkinnedModel model;
+    model.skeleton   = LoadSkeleton(scene);
+    model.animations = LoadAnimations(scene, model.skeleton);
+
+    struct StackItem { const aiNode* node; glm::mat4 parentWorld; };
+    std::vector<StackItem> stack{{ scene->mRootNode, glm::mat4(1.f) }};
+
+    while (!stack.empty()) {
+        auto [node, parentWorld] = stack.back();
+        stack.pop_back();
+
+        // We still traverse with world transforms so non-skinned siblings
+        // work correctly, but skinned meshes are loaded with identity below.
+        const glm::mat4 world = parentWorld * ToGLM(node->mTransformation);
+
+        for (unsigned int c = 0; c < node->mNumChildren; ++c)
+            stack.push_back({ node->mChildren[c], world });
+
+        for (unsigned int m = 0; m < node->mNumMeshes; ++m) {
+            const aiMesh* aiMesh = scene->mMeshes[node->mMeshes[m]];
+            std::string name = aiMesh->mName.length > 0
+                ? std::string(aiMesh->mName.C_Str())
+                : ("mesh_" + std::to_string(node->mMeshes[m]));
+
+            // FIX: pass identity matrix for skinned meshes so vertices stay in
+            // bind-pose space. The inverse bind matrices and joint transforms
+            // are already expressed in that space — baking the node world
+            // transform into positions would offset everything incorrectly.
+            const bool isSkinned = aiMesh->HasBones() && !model.skeleton.joints.empty();
+            const glm::mat4 meshWorld = isSkinned ? glm::mat4{1.f} : world;
+
+            CPUMesh cpu = LoadAiMeshIntoCPUMesh(aiMesh, name, meshWorld);
+            LoadSkinningData(cpu, aiMesh, model.skeleton);
+            cpu.skeleton = model.skeleton;
+            model.meshes.push_back(std::move(cpu));
+        }
+    }
+
+    return model;
+}
+
 } // namespace ModelLoader

-#endif // MODELLOADER_H
+#endif // MODELLOADER_H
--- a/destrum/src/Components/Animator.cpp
+++ b/destrum/src/Components/Animator.cpp
@@ -0,0 +1,189 @@
+#include <destrum/Components/Animator.h>
+#include <destrum/Graphics/Pipelines/SkinningPipeline.h>
+#include <destrum/ObjectModel/GameObject.h>
+
+#include <glm/gtc/matrix_transform.hpp>
+#include <glm/gtc/quaternion.hpp>
+
+#include "spdlog/spdlog.h"
+
+Animator::Animator(GameObject& parent)
+    : Component(parent, "Animator") {}
+
+// ─── Component interface ──────────────────────────────────────────────────────
+
+void Animator::Update() {
+    if (!m_current.clip) return;
+
+    // Time delta comes from your engine's time system
+    // const float dt = Time::GetDeltaTime();
+
+    const float dt = 0.016f; // ~60 FPS
+
+    m_current.time += dt * m_current.speed;
+    if (m_current.clip->looped)
+        m_current.time = std::fmod(m_current.time, m_current.clip->duration);
+    else
+        m_current.time = std::min(m_current.time, m_current.clip->duration);
+
+    if (m_previous.clip) {
+        m_previous.time += dt * m_previous.speed;
+        if (m_previous.clip->looped)
+            m_previous.time = std::fmod(m_previous.time, m_previous.clip->duration);
+
+        m_blendT += dt / m_blendDuration;
+        if (m_blendT >= 1.f) {
+            m_blendT   = 1.f;
+            m_previous = {};
+        }
+    }
+
+    spdlog::info("Playing '{}': time = {:.2f}s, blend = {:.2f}",
+                 m_currentClipName.empty() ? "(none)" : m_currentClipName.c_str(),
+                 m_current.time,
+                 m_blendT);
+}
+
+void Animator::ImGuiInspector() {
+    // ImGui::Text("Clip:  %s", m_currentClipName.empty() ? "(none)" : m_currentClipName.c_str());
+    // ImGui::Text("Time:  %.2f", m_current.time);
+    // ImGui::Text("Blend: %.2f", m_blendT);
+    //
+    // if (!isPlaying()) {
+    //     ImGui::BeginDisabled();
+    //     ImGui::Button("Stop");
+    //     ImGui::EndDisabled();
+    // } else if (ImGui::Button("Stop")) {
+    //     stop();
+    // }
+    //
+    // ImGui::Separator();
+    // ImGui::Text("Clips:");
+    // for (auto& [name, _] : m_clips) {
+    //     if (ImGui::Selectable(name.c_str(), name == m_currentClipName))
+    //         play(name);
+    // }
+}
+
+// ─── Animation control ────────────────────────────────────────────────────────
+
+void Animator::addClip(std::shared_ptr<SkeletalAnimation> clip) {
+    m_clips[clip->name] = std::move(clip);
+}
+
+void Animator::play(const std::string& name, float blendTime) {
+    auto it = m_clips.find(name);
+    if (it == m_clips.end()) return;
+
+    if (m_current.clip && blendTime > 0.f) {
+        m_previous      = m_current;
+        m_blendT        = 0.f;
+        m_blendDuration = blendTime;
+    } else {
+        m_previous = {};
+        m_blendT   = 1.f;
+    }
+
+    m_current        = { it->second.get(), 0.f, 1.f };
+    m_currentClipName = name;
+}
+
+void Animator::stop() {
+    m_current         = {};
+    m_previous        = {};
+    m_currentClipName = {};
+}
+
+// ─── Joint matrix upload ──────────────────────────────────────────────────────
+
+std::size_t Animator::uploadJointMatrices(SkinningPipeline& pipeline,
+                                           const Skeleton& skeleton,
+                                           std::size_t frameIndex) {
+    auto matrices = computeJointMatrices(skeleton);
+    return pipeline.appendJointMatrices(matrices, frameIndex);
+}
+
+// ─── Private: pose evaluation ─────────────────────────────────────────────────
+
+std::vector<glm::mat4> Animator::computeJointMatrices(const Skeleton& skeleton) {
+    const std::size_t numJoints = skeleton.joints.size();
+    std::vector<glm::mat4> global(numJoints, glm::mat4{1.f});
+    std::vector<glm::mat4> result(numJoints, glm::mat4{1.f});
+
+    auto findTrack = [](const SkeletalAnimation* clip,
+                        std::uint32_t idx) -> const SkeletalAnimation::Track* {
+        if (!clip) return nullptr;
+        for (auto& t : clip->tracks)
+            if (t.jointIndex == idx) return &t;
+        return nullptr;
+    };
+
+    for (std::uint32_t i = 0; i < numJoints; ++i) {
+        glm::vec3 tr  = {0.f, 0.f, 0.f};
+        glm::quat rot = glm::identity<glm::quat>();
+        glm::vec3 sc  = {1.f, 1.f, 1.f};
+
+        if (const auto* track = findTrack(m_current.clip, i)) {
+            tr  = sampleTranslation(*track, m_current.time);
+            rot = sampleRotation   (*track, m_current.time);
+            sc  = sampleScale      (*track, m_current.time);
+        }
+
+        if (m_previous.clip && m_blendT < 1.f) {
+            if (const auto* prev = findTrack(m_previous.clip, i)) {
+                tr  = glm::mix  (sampleTranslation(*prev, m_previous.time), tr,  m_blendT);
+                rot = glm::slerp(sampleRotation   (*prev, m_previous.time), rot, m_blendT);
+                sc  = glm::mix  (sampleScale      (*prev, m_previous.time), sc,  m_blendT);
+            }
+        }
+
+        glm::mat4 local = glm::translate(glm::mat4{1.f}, tr)
+                        * glm::mat4_cast(rot)
+                        * glm::scale(glm::mat4{1.f}, sc);
+
+        const int parent = skeleton.parentIndex[i];
+        global[i] = (parent < 0) ? local : global[parent] * local;
+
+        result[i] = global[i] * skeleton.inverseBindMatrices[i];
+    }
+
+    return result;
+}
+
+// ─── Keyframe sampling ────────────────────────────────────────────────────────
+
+glm::vec3 Animator::sampleTranslation(const SkeletalAnimation::Track& track, float t) {
+    const auto& kf = track.keyframes;
+    if (kf.size() == 1) return kf[0].translation;
+    for (std::size_t i = 0; i + 1 < kf.size(); ++i) {
+        if (t <= kf[i + 1].time) {
+            float f = (t - kf[i].time) / (kf[i + 1].time - kf[i].time);
+            return glm::mix(kf[i].translation, kf[i + 1].translation, f);
+        }
+    }
+    return kf.back().translation;
+}
+
+glm::quat Animator::sampleRotation(const SkeletalAnimation::Track& track, float t) {
+    const auto& kf = track.keyframes;
+    if (kf.size() == 1) return kf[0].rotation;
+    for (std::size_t i = 0; i + 1 < kf.size(); ++i) {
+        if (t <= kf[i + 1].time) {
+            float f = (t - kf[i].time) / (kf[i + 1].time - kf[i].time);
+            return glm::slerp(kf[i].rotation, kf[i + 1].rotation, f);
+        }
+    }
+    return kf.back().rotation;
+}
+
+glm::vec3 Animator::sampleScale(const SkeletalAnimation::Track& track, float t) {
+    const auto& kf = track.keyframes;
+    if (kf.size() == 1) return kf[0].scale;
+    for (std::size_t i = 0; i + 1 < kf.size(); ++i) {
+        if (t <= kf[i + 1].time) {
+            float f = (t - kf[i].time) / (kf[i + 1].time - kf[i].time);
+            return glm::mix(kf[i].scale, kf[i + 1].scale, f);
+        }
+    }
+    return kf.back().scale;
+}
--- a/destrum/src/Components/MeshRendererComponent.cpp
+++ b/destrum/src/Components/MeshRendererComponent.cpp
@@ -1,19 +1,56 @@
 #include <destrum/Components/MeshRendererComponent.h>
 #include <destrum/ObjectModel/Transform.h>

+#include "destrum/Components/Animator.h"
+#include "destrum/ObjectModel/GameObject.h"
+#include "destrum/Util/GameState.h"
+

 MeshRendererComponent::MeshRendererComponent(GameObject& parent): Component(parent, "MeshRendererComponent") {
+
 }

 void MeshRendererComponent::Start() {
    Component::Start();
+    if (auto* animator = GetGameObject()->GetComponent<Animator>()) {
+        const auto& gfxDevice = GameState::GetInstance().Gfx();
+        const auto& mesh = GameState::GetInstance().Renderer().GetMeshCache().getMesh(meshID);
+
+        m_skinnedMesh = std::make_unique<SkinnedMesh>();
+        m_skinnedMesh->skinnedVertexBuffer = gfxDevice.createBuffer(
+            mesh.numVertices * sizeof(CPUMesh::Vertex),
+            VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
+            VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT |
+            VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
+        );
+    }
 }

 void MeshRendererComponent::Update() {
 }

 void MeshRendererComponent::Render(const RenderContext& ctx) {
-    if (meshID != NULL_MESH_ID && materialID != NULL_MATERIAL_ID) {
+    if (meshID == NULL_MESH_ID || materialID == NULL_MATERIAL_ID) return;
+
+    if (auto* animator = GetGameObject()->GetComponent<Animator>(); animator && m_skinnedMesh) {
+        const auto& mesh = ctx.renderer.GetMeshCache().getCPUMesh(meshID);
+        const auto skeleton = GetGameObject()->GetComponent<Animator>()->getSkeleton();
+        std::uint32_t frameIdx = GameState::GetInstance().Gfx().getCurrentFrameIndex();
+
+        const std::size_t jointMatricesStartIndex = animator->uploadJointMatrices(
+            ctx.renderer.getSkinningPipeline(),
+            *skeleton, // skeleton stored on GPUMesh (or pass from CPUMesh)
+            frameIdx
+        );
+
+        ctx.renderer.drawSkinnedMesh(
+            meshID,
+            GetTransform().GetWorldMatrix(),
+            materialID,
+            m_skinnedMesh.get(),
+            jointMatricesStartIndex
+        );
+    } else {
        ctx.renderer.drawMesh(meshID, GetTransform().GetWorldMatrix(), materialID);
    }
 }
--- a/destrum/src/Components/OrbitAndSpin.cpp
+++ b/destrum/src/Components/OrbitAndSpin.cpp
@@ -98,13 +98,13 @@ void OrbitAndSpin::Update()
    // GetTransform().SetLocalScale(glm::vec3(std::sin(m_GrowPhase)));

    // material color
-    auto& mat = GameState::GetInstance().Renderer().getMaterialMutable(m_MaterialID);
-    mat.baseColor = glm::vec3(
-        0.5f + 0.5f * std::sin(m_OrbitAngle * 2.0f),
-        0.5f + 0.5f * std::sin(m_OrbitAngle * 3.0f + 2.0f),
-        0.5f + 0.5f * std::sin(m_OrbitAngle * 4.0f + 4.0f)
-    );
-    GameState::GetInstance().Renderer().updateMaterialGPU(m_MaterialID);
+    // auto& mat = GameState::GetInstance().Renderer().getMaterialMutable(m_MaterialID);
+    // mat.baseColor = glm::vec3(
+    //     0.5f + 0.5f * std::sin(m_OrbitAngle * 2.0f),
+    //     0.5f + 0.5f * std::sin(m_OrbitAngle * 3.0f + 2.0f),
+    //     0.5f + 0.5f * std::sin(m_OrbitAngle * 4.0f + 4.0f)
+    // );
+    // GameState::GetInstance().Renderer().updateMaterialGPU(m_MaterialID);
 }

 void OrbitAndSpin::Start() {
--- a/destrum/src/Graphics/BindlessSetManager.cpp
+++ b/destrum/src/Graphics/BindlessSetManager.cpp
@@ -7,11 +7,11 @@

 namespace
 {
-static const std::uint32_t maxBindlessResources = 16536;
-static const std::uint32_t maxSamplers = 32;
+    constexpr std::uint32_t maxBindlessResources = 16536;
+    constexpr std::uint32_t maxSamplers = 32;

-static const std::uint32_t texturesBinding = 0;
-static const std::uint32_t samplersBinding = 1;
+    constexpr std::uint32_t texturesBinding = 0;
+    constexpr std::uint32_t samplersBinding = 1;
 }

 void BindlessSetManager::init(VkDevice device, float maxAnisotropy)
@@ -26,7 +26,7 @@ void BindlessSetManager::init(VkDevice device, float maxAnisotropy)
            .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
            .flags = VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT_EXT,
            .maxSets = 10,
-            .poolSizeCount = (std::uint32_t)poolSizesBindless.size(),
+            .poolSizeCount = static_cast<std::uint32_t>(poolSizesBindless.size()),
            .pPoolSizes = poolSizesBindless.data(),
        };

--- a/destrum/src/Graphics/ComputePipeline.cpp
+++ b/destrum/src/Graphics/ComputePipeline.cpp
@@ -0,0 +1,105 @@
+#include <destrum/Graphics/ComputePipeline.h>
+
+#include <cassert>
+#include <filesystem>
+#include <fstream>
+#include <iostream>
+
+#include "destrum/Util/GameState.h"
+#include "spdlog/spdlog.h"
+
+ComputePipeline::ComputePipeline(GfxDevice& device,
+                                 const std::string& compPath,
+                                 const ComputePipelineConfigInfo& configInfo)
+    : m_device(device) {
+    CreateComputePipeline(compPath, configInfo);
+}
+
+ComputePipeline::~ComputePipeline() {
+    if (m_compShaderModule != VK_NULL_HANDLE) {
+        vkDestroyShaderModule(m_device.getDevice(), m_compShaderModule, nullptr);
+    }
+    if (m_computePipeline != VK_NULL_HANDLE) {
+        vkDestroyPipeline(m_device.getDevice(), m_computePipeline, nullptr);
+    }
+}
+
+void ComputePipeline::bind(VkCommandBuffer buffer) const {
+    vkCmdBindPipeline(buffer, VK_PIPELINE_BIND_POINT_COMPUTE, m_computePipeline);
+}
+
+void ComputePipeline::DefaultPipelineConfigInfo(ComputePipelineConfigInfo& configInfo) {
+    configInfo.name = "DefaultComputePipelineConfigInfo";
+    configInfo.specializationInfo = nullptr;
+    configInfo.pipelineLayout = VK_NULL_HANDLE;
+}
+
+std::vector<char> ComputePipeline::readFile(const std::string& filename) {
+    std::ifstream file(filename, std::ios::ate | std::ios::binary);
+
+    if (!file.is_open()) {
+        throw std::runtime_error("Failed to open file: " + filename);
+    }
+
+    const size_t fileSize = static_cast<size_t>(file.tellg());
+    std::vector<char> buffer(fileSize);
+
+    file.seekg(0);
+    file.read(buffer.data(), static_cast<std::streamsize>(fileSize));
+    file.close();
+
+    return buffer;
+}
+
+void ComputePipeline::CreateComputePipeline(const std::string& compPath,
+                                            const ComputePipelineConfigInfo& configInfo) {
+    assert(configInfo.pipelineLayout != VK_NULL_HANDLE && "no pipelineLayout provided in configInfo");
+    assert(!compPath.empty() && "Compute shader path is empty");
+
+    const std::string compFileName = std::filesystem::path(compPath).filename().string();
+    auto compCode = readFile(compPath);
+
+    spdlog::debug("Compute shader code size: {}", compCode.size());
+    spdlog::debug("Compute shader file: {}", compFileName);
+
+    CreateShaderModule(compCode, &m_compShaderModule);
+
+    VkPipelineShaderStageCreateInfo compStageInfo{};
+    compStageInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
+    compStageInfo.stage = VK_SHADER_STAGE_COMPUTE_BIT;
+    compStageInfo.module = m_compShaderModule;
+    compStageInfo.pName = "main";
+    compStageInfo.pSpecializationInfo = configInfo.specializationInfo;
+
+    VkComputePipelineCreateInfo pipelineInfo{};
+    pipelineInfo.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO;
+    pipelineInfo.stage = compStageInfo;
+    pipelineInfo.layout = configInfo.pipelineLayout;
+
+    pipelineInfo.basePipelineIndex = -1;
+    pipelineInfo.basePipelineHandle = VK_NULL_HANDLE;
+
+    if (vkCreateComputePipelines(m_device.getDevice(),
+                                 VK_NULL_HANDLE,
+                                 1,
+                                 &pipelineInfo,
+                                 nullptr,
+                                 &m_computePipeline) != VK_SUCCESS) {
+        throw std::runtime_error("Can't make compute pipeline!");
+    }
+
+    if (!configInfo.name.empty()) {
+        vkutil::addDebugLabel(GameState::GetInstance().Gfx().getDevice(), m_computePipeline, configInfo.name.c_str());
+    }
+}
+
+void ComputePipeline::CreateShaderModule(const std::vector<char>& code, VkShaderModule* shaderModule) const {
+    VkShaderModuleCreateInfo createInfo{};
+    createInfo.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO;
+    createInfo.codeSize = code.size();
+    createInfo.pCode = reinterpret_cast<const uint32_t*>(code.data());
+
+    if (vkCreateShaderModule(m_device.getDevice(), &createInfo, nullptr, shaderModule) != VK_SUCCESS) {
+        throw std::runtime_error("Failed to create shader module!");
+    }
+}
--- a/destrum/src/Graphics/Frustum.cpp
+++ b/destrum/src/Graphics/Frustum.cpp
@@ -0,0 +1,128 @@
+#include <destrum/Graphics/Frustum.h>
+
+Frustum edge::createFrustumFromCamera(const Camera& camera) {
+// TODO: write a non-horrible version of this, lol
+    Frustum frustum;
+//     if (camera.isOrthographic()) {
+//         const auto points = calculateFrustumCornersWorldSpace(camera);
+//
+//         /*
+//                5────────6
+//               ╱┊       ╱│
+//              ╱ ┊      ╱ │
+//             1──┼─────2  │
+//             │  ┊ (C) │  │            Y ╿   . Z
+//             │  4┈┈┈┈┈│┈┈7              │  ╱
+//             │ ╱      │ ╱         X     │ ╱
+//             │╱       │╱           ╾────┼
+//             0--------3
+//
+//         */
+//         // from bottom-left and moving CW...
+//         static const std::array<int, 4> near{0, 1, 2, 3};
+//         static const std::array<int, 4> far{7, 6, 5, 4};
+//         static const std::array<int, 4> left{4, 5, 1, 0};
+//         static const std::array<int, 4> right{3, 2, 6, 7};
+//         static const std::array<int, 4> bottom{4, 0, 3, 7};
+//         static const std::array<int, 4> top{5, 6, 2, 1};
+//
+//         frustum.nearFace = {findCenter(points, near), findNormal(points, near)};
+//         frustum.farFace = {findCenter(points, far), findNormal(points, far)};
+//         frustum.leftFace = {findCenter(points, left), findNormal(points, left)};
+//         frustum.rightFace = {findCenter(points, right), findNormal(points, right)};
+//         frustum.bottomFace = {findCenter(points, bottom), findNormal(points, bottom)};
+//         frustum.topFace = {findCenter(points, top), findNormal(points, top)};
+//     } else {
+        const auto camPos = camera.GetPosition();
+        const auto camFront = camera.GetForward();
+        const auto camUp = camera.GetUp();
+        const auto camRight = camera.GetRight();
+
+        const auto zNear = camera.GetZNear();
+        const auto zFar = camera.GetZFar();
+        const auto halfVSide = zFar * tanf(camera.GetFOVY() * .5f);
+        const auto halfHSide = halfVSide * camera.GetAspectRatio();
+        const auto frontMultFar = zFar * camFront;
+
+        frustum.nearFace = {camPos + zNear * camFront, camFront};
+        frustum.farFace = {camPos + frontMultFar, -camFront};
+        frustum.leftFace = {camPos, glm::cross(camUp, frontMultFar + camRight * halfHSide)};
+        frustum.rightFace = {camPos, glm::cross(frontMultFar - camRight * halfHSide, camUp)};
+        frustum.bottomFace = {camPos, glm::cross(frontMultFar + camUp * halfVSide, camRight)};
+        frustum.topFace = {camPos, glm::cross(camRight, frontMultFar - camUp * halfVSide)};
+    // }
+
+    return frustum;
+}
+
+bool isOnOrForwardPlane(const Frustum::Plane& plane, const Sphere& sphere)
+{
+    return plane.getSignedDistanceToPlane(sphere.center) > -sphere.radius;
+}
+
+bool edge::isInFrustum(const Frustum& frustum, const Sphere& s) {
+    return (
+       isOnOrForwardPlane(frustum.farFace, s) && isOnOrForwardPlane(frustum.nearFace, s) &&
+       isOnOrForwardPlane(frustum.leftFace, s) && isOnOrForwardPlane(frustum.rightFace, s) &&
+       isOnOrForwardPlane(frustum.topFace, s) && isOnOrForwardPlane(frustum.bottomFace, s));
+}
+
+bool edge::isInFrustum(const Frustum& frustum, const AABB& aabb) {
+    glm::vec3 vmin, vmax;
+    bool ret = true;
+    for (int i = 0; i < 6; ++i) {
+        const auto& plane = frustum.getPlane(i);
+        // X axis
+        if (plane.normal.x < 0) {
+            vmin.x = aabb.min.x;
+            vmax.x = aabb.max.x;
+        } else {
+            vmin.x = aabb.max.x;
+            vmax.x = aabb.min.x;
+        }
+        // Y axis
+        if (plane.normal.y < 0) {
+            vmin.y = aabb.min.y;
+            vmax.y = aabb.max.y;
+        } else {
+            vmin.y = aabb.max.y;
+            vmax.y = aabb.min.y;
+        }
+        // Z axis
+        if (plane.normal.z < 0) {
+            vmin.z = aabb.min.z;
+            vmax.z = aabb.max.z;
+        } else {
+            vmin.z = aabb.max.z;
+            vmax.z = aabb.min.z;
+        }
+        if (plane.getSignedDistanceToPlane(vmin) < 0) {
+            return false;
+        }
+        if (plane.getSignedDistanceToPlane(vmax) <= 0) {
+            ret = true;
+        }
+    }
+    return ret;
+}
+
+glm::vec3 getTransformScale(const glm::mat4& transform)
+{
+    float sx = glm::length(glm::vec3{transform[0][0], transform[0][1], transform[0][2]});
+    float sy = glm::length(glm::vec3{transform[1][0], transform[1][1], transform[1][2]});
+    float sz = glm::length(glm::vec3{transform[2][0], transform[2][1], transform[2][2]});
+    return {sx, sy, sz};
+}
+
+Sphere edge::calculateBoundingSphereWorld(const glm::mat4& transform, const Sphere& s, bool hasSkeleton) {
+    const auto scale = getTransformScale(transform);
+    float maxScale = std::max({scale.x, scale.y, scale.z});
+    if (hasSkeleton) {
+        maxScale = 5.f; // ignore scale for skeleton meshes (TODO: fix)
+        // setting scale to 1.f causes prolems with frustum culling
+    }
+    auto sphereWorld = s;
+    sphereWorld.radius *= maxScale;
+    sphereWorld.center = glm::vec3(transform * glm::vec4(sphereWorld.center, 1.f));
+    return sphereWorld;
+}
--- a/destrum/src/Graphics/GfxDevice.cpp
+++ b/destrum/src/Graphics/GfxDevice.cpp
@@ -129,18 +129,36 @@ void GfxDevice::init(SDL_Window* window, const std::string& appName, bool vSync)
        auto& mainCommandBuffer = frames[i].commandBuffer;
        VK_CHECK(vkAllocateCommandBuffers(device, &cmdAllocInfo, &mainCommandBuffer));
    }
-    //
-    // { // create white texture
-    //     std::uint32_t pixel = 0xFFFFFFFF;
-    //     whiteImageId = createImage(
-    //         {
-    //             .format = VK_FORMAT_R8G8B8A8_UNORM,
-    //             .usage = VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT,
-    //             .extent = VkExtent3D{1, 1, 1},
-    //         },
-    //         "white texture",
-    //         &pixel);
-    // }
+
+    { // create white texture
+        std::uint32_t pixel = 0xFFFFFFFF;
+        whiteImageId = createImage(
+            {
+                .format = VK_FORMAT_R8G8B8A8_UNORM,
+                .usage = VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT,
+                .extent = VkExtent3D{1, 1, 1},
+            },
+            "white texture",
+            &pixel);
+    }
+
+    { // create error texture (black/magenta checker)
+        constexpr auto black = 0xFF000000;
+        constexpr auto magenta = 0xFFFF00FF;
+
+        std::array<std::uint32_t, 4> pixels{black, magenta, magenta, black};
+        errorImageId = createImage(
+            {
+                .format = VK_FORMAT_R8G8B8A8_UNORM,
+                .usage = VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT |
+                         VK_IMAGE_USAGE_TRANSFER_SRC_BIT,
+                .extent = VkExtent3D{2, 2, 1},
+            },
+            "error texture",
+            pixels.data());
+        imageCache.setErrorImageId(errorImageId);
+    }
+
    GameState::GetInstance().SetGfxDevice(this);
 }

@@ -495,25 +513,14 @@ GPUImage GfxDevice::createImageRaw(
    return image;
 }

-GPUImage GfxDevice::loadImageFromFileRaw(
-    const std::filesystem::path& path,
-    VkImageUsageFlags usage,
-    bool mipMap,
-    TextureIntent intent) const {
-    // 1) Decode file to CPU pixels (stb for png/jpg/hdr, tinyexr for exr)
-    //    IMPORTANT: util::loadImage should fill:
-    //      - width/height/channels(=4)
-    //      - hdr + (pixels OR hdrPixels)
-    //      - vkFormat (RGBA8_SRGB or RGBA8_UNORM or RGBA32F)
-    //      - byteSize (exact bytes to upload)
+GPUImage GfxDevice::loadImageFromFileRaw(const std::filesystem::path& path, VkImageUsageFlags usage, bool mipMap, TextureIntent intent) const {
    const auto data = util::loadImage(path, intent);
-    if (data.vkFormat == VK_FORMAT_UNDEFINED || data.byteSize == 0 ||
-        (data.hdr ? (data.hdrPixels == nullptr) : (data.pixels == nullptr))) {
+
+    if (data.vkFormat == VK_FORMAT_UNDEFINED || data.byteSize == 0 || (data.hdr ? (data.hdrPixels == nullptr) : (data.pixels == nullptr))) {
        spdlog::error("Failed to load image '{}'", path.string());
        return getImage(errorImageId);
    }

-    // 2) Create GPU image using the format the loader chose (matches CPU memory layout)
    auto image = createImageRaw({
        .format = data.vkFormat,
        .usage = usage |
@@ -527,15 +534,10 @@ GPUImage GfxDevice::loadImageFromFileRaw(
        .mipMap = mipMap,
    });

-    // 3) Upload *exactly* byteSize bytes from the correct pointer
-    const void* src = data.hdr
-                          ? static_cast<const void*>(data.hdrPixels)
-                          : static_cast<const void*>(data.pixels);
+    const void* src = data.hdr ? static_cast<const void*>(data.hdrPixels) : static_cast<const void*>(data.pixels);

-    // Use the "sized" upload to avoid BytesPerTexel mismatches
    uploadImageDataSized(image, src, data.byteSize, 0);

-    // 4) Debug label
    image.debugName = path.string();
    vkutil::addDebugLabel(device, image.image, path.string().c_str());

--- a/destrum/src/Graphics/ImageCache.cpp
+++ b/destrum/src/Graphics/ImageCache.cpp
@@ -2,21 +2,12 @@

 #include <destrum/Graphics/GfxDevice.h>

-ImageCache::ImageCache(GfxDevice& gfxDevice) : gfxDevice(gfxDevice)
-{}
+ImageCache::ImageCache(GfxDevice& gfxDevice) : gfxDevice(gfxDevice) {
+}

-ImageID ImageCache::loadImageFromFile(
-    const std::filesystem::path& path,
-    VkImageUsageFlags usage,
-    bool mipMap,
-    TextureIntent intent)
-{
-    for (const auto& [id, info] : loadedImagesInfo) {
-        if (info.path == path &&
-            info.intent == intent &&
-            info.usage == usage &&
-            info.mipMap == mipMap)
-        {
+ImageID ImageCache::loadImageFromFile(const std::filesystem::path& path, VkImageUsageFlags usage, bool mipMap, TextureIntent intent) {
+    for (const auto& [id, info]: loadedImagesInfo) {
+        if (info.path == path && info.intent == intent && info.usage == usage && info.mipMap == mipMap) {
            return id;
        }
    }
@@ -41,13 +32,11 @@ ImageID ImageCache::loadImageFromFile(
    return id;
 }

-ImageID ImageCache::addImage(GPUImage image)
-{
+ImageID ImageCache::addImage(GPUImage image) {
    return addImage(getFreeImageId(), std::move(image));
 }

-ImageID ImageCache::addImage(ImageID id, GPUImage image)
-{
+ImageID ImageCache::addImage(ImageID id, GPUImage image) {
    image.setBindlessId(static_cast<std::uint32_t>(id));
    if (id != images.size()) {
        images[id] = std::move(image); // replacing existing image
@@ -59,20 +48,17 @@ ImageID ImageCache::addImage(ImageID id, GPUImage image)
    return id;
 }

-const GPUImage& ImageCache::getImage(ImageID id) const
-{
+const GPUImage& ImageCache::getImage(ImageID id) const {
    assert(id != NULL_IMAGE_ID && id < images.size());
    return images.at(id);
 }

-ImageID ImageCache::getFreeImageId() const
-{
+ImageID ImageCache::getFreeImageId() const {
    return images.size();
 }

-void ImageCache::destroyImages()
-{
-    for (const auto& image : images) {
+void ImageCache::destroyImages() {
+    for (const auto& image: images) {
        gfxDevice.destroyImage(image);
    }
    images.clear();
--- a/destrum/src/Graphics/ImageLoader.cpp
+++ b/destrum/src/Graphics/ImageLoader.cpp
@@ -37,6 +37,8 @@ namespace util
    {
        ImageData data;

+
+
        // ---------- EXR ----------
        if (isExrExt(p)) {
            float* out = nullptr;
--- a/destrum/src/Graphics/MaterialCache.cpp
+++ b/destrum/src/Graphics/MaterialCache.cpp
@@ -24,7 +24,7 @@ void MaterialCache::init(GfxDevice& gfxDevice)
            &normal);
    }

-    Material placeholderMaterial{.name = "PLACEHOLDER_MATERIAL"};
+    Material placeholderMaterial{.diffuseTexture = defaultNormalMapTextureID, .name = "PLACEHOLDER_MATERIAL"};
    placeholderMaterialId = addMaterial(gfxDevice, placeholderMaterial);
 }

@@ -41,13 +41,14 @@ MaterialID MaterialCache::addMaterial(GfxDevice& gfxDevice, Material material)
    };

    // store on GPU
-    MaterialData* data = (MaterialData*)materialDataBuffer.info.pMappedData;
-    auto whiteTextureID = gfxDevice.getWhiteTextureID();
-    auto id = getFreeMaterialId();
+    MaterialData* data = static_cast<MaterialData*>(materialDataBuffer.info.pMappedData);
+    const auto whiteTextureID = gfxDevice.getWhiteTextureID();
+    const auto id = getFreeMaterialId();
    assert(id < MAX_MATERIALS);
    data[id] = MaterialData{
        .baseColor = glm::vec4(material.baseColor, 1.0f),
        .metalRoughnessEmissive = glm::vec4{material.metallicFactor, material.roughnessFactor, material.emissiveFactor, 0.f},
+        .textureFilteringMode = static_cast<std::uint32_t>(material.textureFilteringMode),
        .diffuseTex = getTextureOrElse(material.diffuseTexture, whiteTextureID),
        .normalTex = whiteTextureID,
        .metallicRoughnessTex = whiteTextureID,
@@ -60,6 +61,15 @@ MaterialID MaterialCache::addMaterial(GfxDevice& gfxDevice, Material material)
    return id;
 }

+MaterialID MaterialCache::addSimpleTextureMaterial(GfxDevice& gfxDevice, ImageID textureID) {
+    Material material{};
+    material.name = "idk";
+    material.diffuseTexture = textureID;
+    material.metallicFactor = 0.0f;
+    material.roughnessFactor = 1.0f;
+    return addMaterial(gfxDevice, material);
+}
+
 const Material& MaterialCache::getMaterial(MaterialID id) const
 {
    return materials.at(id);
@@ -100,7 +110,7 @@ void MaterialCache::updateMaterialGPU(GfxDevice& gfxDevice, MaterialID id)
        .baseColor = glm::vec4(material.baseColor, 1.0f),
        .metalRoughnessEmissive = glm::vec4(material.metallicFactor, material.roughnessFactor, material.emissiveFactor, 0.f),
        .diffuseTex           = getTextureOrElse(material.diffuseTexture, whiteTextureID),
-        .normalTex            = whiteTextureID, // if you have this field
+        .normalTex            = whiteTextureID,
        .metallicRoughnessTex = whiteTextureID,
        .emissiveTex          = whiteTextureID,
    };
--- a/destrum/src/Graphics/MeshCache.cpp
+++ b/destrum/src/Graphics/MeshCache.cpp
@@ -3,25 +3,30 @@
 #include <destrum/Graphics/Resources/Mesh.h>
 #include <destrum/Graphics/GfxDevice.h>
 #include <destrum/Graphics/Util.h>
+
+#include <destrum/Util/MathUtils.h>
 // #include <destrum/Math/Util.h>

 MeshID MeshCache::addMesh(GfxDevice& gfxDevice, const CPUMesh& cpuMesh)
 {
    auto gpuMesh = GPUMesh{
        .numVertices = (std::uint32_t)cpuMesh.vertices.size(),
-        .numIndices = (std::uint32_t)cpuMesh.indices.size(),
-        .minPos = cpuMesh.minPos,
-        .maxPos = cpuMesh.maxPos,
+        .numIndices  = (std::uint32_t)cpuMesh.indices.size(),
+        .minPos      = cpuMesh.minPos,
+        .maxPos      = cpuMesh.maxPos,
+        .hasSkeleton = cpuMesh.skeleton.has_value(),
    };

    std::vector<glm::vec3> positions(cpuMesh.vertices.size());
-    for (std::size_t i = 0; i < cpuMesh.vertices.size(); ++i) {
+    for (std::size_t i = 0; i < cpuMesh.vertices.size(); ++i)
        positions[i] = cpuMesh.vertices[i].position;
-    }
+    gpuMesh.boundingSphere = calculateBoundingSphere(positions);

    uploadMesh(gfxDevice, cpuMesh, gpuMesh);
+
    const auto id = meshes.size();
    meshes.push_back(std::move(gpuMesh));
+    cpuMeshes.push_back(cpuMesh);        // store a copy of the CPU mesh
    return id;
 }

@@ -70,6 +75,33 @@ void MeshCache::uploadMesh(GfxDevice& gfxDevice, const CPUMesh& cpuMesh, GPUMesh
    const auto idxBufferName = cpuMesh.name + " (idx)";
    vkutil::addDebugLabel(gfxDevice.getDevice(), gpuMesh.vertexBuffer.buffer, vtxBufferName.c_str());
    vkutil::addDebugLabel(gfxDevice.getDevice(), gpuMesh.indexBuffer.buffer, idxBufferName.c_str());
+
+    if (gpuMesh.hasSkeleton) {
+        // create skinning data buffer
+        const auto skinningDataSize = cpuMesh.vertices.size() * sizeof(CPUMesh::SkinningData);
+        gpuMesh.skinningDataBuffer = gfxDevice.createBuffer(
+            skinningDataSize,
+            VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT |
+                VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT);
+
+        const auto staging =
+            gfxDevice.createBuffer(skinningDataSize, VK_BUFFER_USAGE_TRANSFER_SRC_BIT);
+
+        // copy data
+        void* data = staging.info.pMappedData;
+        memcpy(data, cpuMesh.skinningData.data(), skinningDataSize);
+
+        gfxDevice.immediateSubmit([&](VkCommandBuffer cmd) {
+            const auto vertexCopy = VkBufferCopy{
+                .srcOffset = 0,
+                .dstOffset = 0,
+                .size = skinningDataSize,
+            };
+            vkCmdCopyBuffer(cmd, staging.buffer, gpuMesh.skinningDataBuffer.buffer, 1, &vertexCopy);
+        });
+
+        gfxDevice.destroyBuffer(staging);
+    }
 }

 const GPUMesh& MeshCache::getMesh(MeshID id) const
@@ -77,6 +109,11 @@ const GPUMesh& MeshCache::getMesh(MeshID id) const
    return meshes.at(id);
 }

+const CPUMesh& MeshCache::getCPUMesh(MeshID id) const
+{
+    return cpuMeshes.at(id);
+}
+
 void MeshCache::cleanup(const GfxDevice& gfxDevice)
 {
    for (const auto& mesh : meshes) {
--- a/destrum/src/Graphics/Pipelines/MeshPipeline.cpp
+++ b/destrum/src/Graphics/Pipelines/MeshPipeline.cpp
@@ -1,6 +1,9 @@
 #include <destrum/Graphics/Pipelines/MeshPipeline.h>
 #include <destrum/FS/AssetFS.h>

+#include "destrum/Graphics/Frustum.h"
+#include "spdlog/spdlog.h"
+
 MeshPipeline::MeshPipeline(): m_pipelineLayout{nullptr} {
 }

@@ -68,6 +71,8 @@ void MeshPipeline::draw(VkCommandBuffer cmd,
    m_pipeline->bind(cmd);
    gfxDevice.bindBindlessDescSet(cmd, m_pipelineLayout);

+    int ActualDrawCalls = 0;
+
    const auto viewport = VkViewport{
        .x = 0,
        .y = 0,
@@ -86,16 +91,17 @@ void MeshPipeline::draw(VkCommandBuffer cmd,

    auto prevMeshId = NULL_MESH_ID;

-    // const auto frustum = edge::createFrustumFromCamera(camera);
+    const auto frustum = edge::createFrustumFromCamera(camera);

    for (const auto& dcIdx : drawCommands) {
-        // const auto& dc = drawCommands[dcIdx];
        const auto& dc = dcIdx;

        // if (!edge::isInFrustum(frustum, dc.worldBoundingSphere)) {
-            // continue;
+        //     continue;
        // }

+        ActualDrawCalls++;
+
        const auto& mesh = meshCache.getMesh(dc.meshId);
        if (dc.meshId != prevMeshId) {
            prevMeshId = dc.meshId;
@@ -106,7 +112,7 @@ void MeshPipeline::draw(VkCommandBuffer cmd,
        const auto pushConstants = PushConstants{
            .transform = dc.transformMatrix,
            .sceneDataBuffer = sceneDataBuffer.address,
-            .vertexBuffer = mesh.vertexBuffer.address,
+            .vertexBuffer = dc.skinnedMesh != nullptr ? dc.skinnedMesh->skinnedVertexBuffer.address : mesh.vertexBuffer.address,
            .materialId = dc.materialId,
        };
        vkCmdPushConstants(
--- a/destrum/src/Graphics/Pipelines/SkinningPipeline.cpp
+++ b/destrum/src/Graphics/Pipelines/SkinningPipeline.cpp
@@ -0,0 +1,98 @@
+#include <destrum/Graphics/Pipelines/SkinningPipeline.h>
+
+#include "destrum/FS/AssetFS.h"
+#include "destrum/Graphics/MeshCache.h"
+#include "destrum/Graphics/MeshDrawCommand.h"
+
+void SkinningPipeline::init(GfxDevice& gfxDevice) {
+    const auto& device = gfxDevice.getDevice();
+
+    const auto pushConstant = VkPushConstantRange{
+        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+        .offset = 0,
+        .size = sizeof(PushConstants),
+    };
+
+    const auto pushConstants = std::array{pushConstant};
+
+    VkPipelineLayoutCreateInfo pipelineLayoutInfo{};
+    pipelineLayoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
+
+    pipelineLayoutInfo.pushConstantRangeCount = 1;
+    pipelineLayoutInfo.pPushConstantRanges = pushConstants.data();
+
+    if (vkCreatePipelineLayout(gfxDevice.getDevice().device, &pipelineLayoutInfo, nullptr, &m_pipelineLayout) != VK_SUCCESS) {
+        throw std::runtime_error("Could not make pipleine layout");
+    }
+
+
+    ComputePipelineConfigInfo pipelineConfig{};
+    pipelineConfig.name = "skinning compute pipeline";
+    pipelineConfig.pipelineLayout = m_pipelineLayout;
+
+    const auto SkinningShaderPath = AssetFS::GetInstance().GetCookedPathForFile("engine://shaders/skinning.comp");
+
+    skinningPipeline = std::make_unique<ComputePipeline>(
+        gfxDevice,
+        SkinningShaderPath.generic_string(),
+        pipelineConfig
+    );
+
+
+    for (std::size_t i = 0; i < FRAMES_IN_FLIGHT; ++i) {
+        auto& jointMatricesBuffer = framesData[i].jointMatricesBuffer;
+        jointMatricesBuffer.capacity = MAX_JOINT_MATRICES;
+        jointMatricesBuffer.buffer = gfxDevice.createBuffer(
+            MAX_JOINT_MATRICES * sizeof(glm::mat4),
+            VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT);
+    }
+}
+
+void SkinningPipeline::cleanup(GfxDevice& gfxDevice) {
+    for (auto& frame : framesData) {
+        gfxDevice.destroyBuffer(frame.jointMatricesBuffer.buffer);
+    }
+    vkDestroyPipelineLayout(gfxDevice.getDevice().device, m_pipelineLayout, nullptr);
+    skinningPipeline.reset();
+}
+
+void SkinningPipeline::doSkinning(VkCommandBuffer cmd, std::size_t frameIndex, const MeshCache& meshCache, const MeshDrawCommand& dc) {
+
+    skinningPipeline->bind(cmd);
+
+    const auto& mesh = meshCache.getMesh(dc.meshId);
+    assert(mesh.hasSkeleton);
+    assert(dc.skinnedMesh);
+
+    const auto cs = PushConstants{
+        .jointMatricesBuffer = getCurrentFrameData(frameIndex).jointMatricesBuffer.buffer.address,
+        .jointMatricesStartIndex = static_cast<std::uint32_t>(dc.jointMatricesStartIndex),  // explicit cast
+        .numVertices = mesh.numVertices,
+        .inputBuffer = mesh.vertexBuffer.address,
+        .skinningData = mesh.skinningDataBuffer.address,
+        .outputBuffer = dc.skinnedMesh->skinnedVertexBuffer.address,
+    };
+    vkCmdPushConstants(cmd, m_pipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(PushConstants), &cs);
+
+    static const auto workgroupSize = 256;
+    // const auto groupSizeX = (std::uint32_t)std::ceil(mesh.numVertices / (float)workgroupSize);
+    const auto groupSizeX = static_cast<std::uint32_t>(
+    std::ceil(mesh.numVertices / (float)workgroupSize));
+    vkCmdDispatch(cmd, groupSizeX, 1, 1);
+}
+
+void SkinningPipeline::beginDrawing(std::size_t frameIndex) {
+    getCurrentFrameData(frameIndex).jointMatricesBuffer.clear();
+}
+
+std::size_t SkinningPipeline::appendJointMatrices(std::span<const glm::mat4> jointMatrices, std::size_t frameIndex) {
+
+    auto& jointMatricesBuffer = getCurrentFrameData(frameIndex).jointMatricesBuffer;
+    const auto startIndex = jointMatricesBuffer.size;
+    jointMatricesBuffer.append(jointMatrices);
+    return startIndex;
+}
+
+SkinningPipeline::PerFrameData& SkinningPipeline::getCurrentFrameData(std::size_t frameIndex) {
+    return framesData[frameIndex % FRAMES_IN_FLIGHT];
+}
--- a/destrum/src/Graphics/Renderer.cpp
+++ b/destrum/src/Graphics/Renderer.cpp
@@ -23,6 +23,9 @@ void GameRenderer::init(GfxDevice& gfxDevice, const glm::ivec2& drawImageSize) {
    skyboxPipeline = std::make_unique<SkyboxPipeline>();
    skyboxPipeline->init(gfxDevice, drawImageFormat, depthImageFormat);

+    skinningPipeline = std::make_unique<SkinningPipeline>();
+    skinningPipeline->init(gfxDevice);
+

    GameState::GetInstance().SetRenderer(this);
 }
@@ -30,6 +33,7 @@ void GameRenderer::init(GfxDevice& gfxDevice, const glm::ivec2& drawImageSize) {
 void GameRenderer::beginDrawing(GfxDevice& gfxDevice) {
    flushMaterialUpdates(gfxDevice);
    meshDrawCommands.clear();
+    skinningPipeline->beginDrawing(gfxDevice.getCurrentFrameIndex());
 }

 void GameRenderer::endDrawing() {
@@ -37,6 +41,13 @@ void GameRenderer::endDrawing() {
 }

 void GameRenderer::draw(VkCommandBuffer cmd, GfxDevice& gfxDevice, const Camera& camera, const SceneData& sceneData) {
+
+    for (const auto& dc : meshDrawCommands) {
+        if (!dc.skinnedMesh) {
+            continue;
+        }
+        skinningPipeline->doSkinning(cmd, gfxDevice.getCurrentFrameIndex(), meshCache, dc);
+    }
    const auto gpuSceneData = GPUSceneData{
        .view = sceneData.camera.GetViewMatrix(),
        .proj = sceneData.camera.GetProjectionMatrix(),
@@ -107,7 +118,7 @@ void GameRenderer::cleanup(VkDevice device) {

 void GameRenderer::drawMesh(MeshID id, const glm::mat4& transform, MaterialID materialId) {
    const auto& mesh = meshCache.getMesh(id);
-    // const auto worldBoundingSphere = edge::calculateBoundingSphereWorld(transform, mesh.boundingSphere, false);
+    const auto worldBoundingSphere = edge::calculateBoundingSphereWorld(transform, mesh.boundingSphere, false);
    assert(materialId != NULL_MATERIAL_ID);

    meshDrawCommands.push_back(MeshDrawCommand{
@@ -117,6 +128,25 @@ void GameRenderer::drawMesh(MeshID id, const glm::mat4& transform, MaterialID ma
    });
 }

+void GameRenderer::drawSkinnedMesh(MeshID id,
+                                    const glm::mat4& transform,
+                                    MaterialID materialId,
+                                    SkinnedMesh* skinnedMesh,
+                                    std::size_t jointMatricesStartIndex) {
+    const auto& mesh = meshCache.getMesh(id);
+    const auto  worldBoundingSphere = edge::calculateBoundingSphereWorld(transform, mesh.boundingSphere, false);
+    assert(materialId != NULL_MATERIAL_ID);
+    assert(skinnedMesh != nullptr);
+
+    meshDrawCommands.push_back(MeshDrawCommand{
+        .meshId                  = id,
+        .transformMatrix         = transform,
+        .materialId              = materialId,
+        .skinnedMesh             = skinnedMesh,
+        .jointMatricesStartIndex = static_cast<std::uint32_t>(jointMatricesStartIndex),
+    });
+}
+
 const GPUImage& GameRenderer::getDrawImage(const GfxDevice& gfx_device) const {
    return gfx_device.getImage(drawImageId);
 }
--- a/destrum/src/Graphics/Util.cpp
+++ b/destrum/src/Graphics/Util.cpp
@@ -122,6 +122,16 @@ void vkutil::addDebugLabel(VkDevice device, VkShaderModule shaderModule, const c
    vkSetDebugUtilsObjectNameEXT(device, &nameInfo);
 }

+void vkutil::addDebugLabel(VkDevice device, VkPipeline pipeline, const char* label) {
+    const auto nameInfo = VkDebugUtilsObjectNameInfoEXT{
+        .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_OBJECT_NAME_INFO_EXT,
+        .objectType = VK_OBJECT_TYPE_PIPELINE,
+        .objectHandle = (std::uint64_t)pipeline,
+        .pObjectName = label,
+    };
+    vkSetDebugUtilsObjectNameEXT(device, &nameInfo);
+}
+
 void vkutil::addDebugLabel(VkDevice device, VkBuffer buffer, const char* label) {
    const auto nameInfo = VkDebugUtilsObjectNameInfoEXT{
        .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_OBJECT_NAME_INFO_EXT,
--- a/destrum/third_party/CMakeLists.txt
+++ b/destrum/third_party/CMakeLists.txt
@@ -1,4 +1,5 @@
 # glm
+set(BUILD_SHARED_LIBS OFF)
 add_subdirectory(glm)

 # stb
@@ -56,3 +57,10 @@ add_subdirectory(tinygltf)

 add_subdirectory(tinyexr)
 target_include_directories(tinyexr PUBLIC "${CMAKE_CURRENT_LIST_DIR}/tinyexr")
+
+
+set(BUILD_SHARED_LIBS OFF CACHE BOOL "" FORCE)
+set(ASSIMP_BUILD_TESTS OFF CACHE BOOL "" FORCE)
+set(ASSIMP_INJECT_DEBUG_POSTFIX OFF CACHE BOOL "" FORCE)
+set(ASSIMP_INSTALL OFF CACHE BOOL "" FORCE)
+add_subdirectory(assimp)
--- a/destrum/third_party/assimp
+++ b/destrum/third_party/assimp
--- a/lightkeeper/CMakeLists.txt
+++ b/lightkeeper/CMakeLists.txt
@@ -22,14 +22,26 @@ target_include_directories(lightkeeper PRIVATE "${CMAKE_CURRENT_LIST_DIR}/includ

 target_link_libraries(lightkeeper PRIVATE destrum::destrum)

-set(ASSETS_SRC_DIR     "${CMAKE_CURRENT_LIST_DIR}/assets_src")
-set(ASSETS_RUNTIME_DIR "${CMAKE_CURRENT_LIST_DIR}/assets_runtime")
-set(OUTPUT_GAME_ASSETS_DIR "${CMAKE_BINARY_DIR}/assets/game")
+#set(ASSETS_SRC_DIR     "${CMAKE_CURRENT_LIST_DIR}/assets_src")
+#set(ASSETS_RUNTIME_DIR "${CMAKE_CURRENT_LIST_DIR}/assets_runtime")
+#set(OUTPUT_GAME_ASSETS_DIR "${CMAKE_BINARY_DIR}/assets/game")
+#
+#add_custom_command(TARGET lightkeeper POST_BUILD
+#        COMMAND ${CMAKE_COMMAND} -E make_directory "${OUTPUT_GAME_ASSETS_DIR}"
+#        COMMAND ${CMAKE_COMMAND} -E rm -rf "${OUTPUT_GAME_ASSETS_DIR}"
+#        COMMAND ${CMAKE_COMMAND} -E create_symlink "${ASSETS_RUNTIME_DIR}" "${OUTPUT_GAME_ASSETS_DIR}"
+#        VERBATIM
+#)
+
+set(ASSETS_SRC_DIR          "${CMAKE_CURRENT_LIST_DIR}/assets_src")
+set(ASSETS_RUNTIME_DIR      "${CMAKE_CURRENT_LIST_DIR}/assets_runtime")
+set(OUTPUT_GAME_ASSETS_DIR  "${CMAKE_CURRENT_BINARY_DIR}/assets/game")

 add_custom_command(TARGET lightkeeper POST_BUILD
-        COMMAND ${CMAKE_COMMAND} -E make_directory "${OUTPUT_GAME_ASSETS_DIR}"
-        COMMAND ${CMAKE_COMMAND} -E rm -rf "${OUTPUT_GAME_ASSETS_DIR}"
-        COMMAND ${CMAKE_COMMAND} -E create_symlink "${ASSETS_RUNTIME_DIR}" "${OUTPUT_GAME_ASSETS_DIR}"
+        COMMAND ${CMAKE_COMMAND} -E make_directory "${CMAKE_CURRENT_BINARY_DIR}/assets"
+
+        COMMAND ${CMAKE_COMMAND} -E rm -rf "${CMAKE_CURRENT_BINARY_DIR}/assets/game"
+        COMMAND ${CMAKE_COMMAND} -E create_symlink "${ASSETS_RUNTIME_DIR}" "${CMAKE_CURRENT_BINARY_DIR}/assets/game"
        VERBATIM
 )

@@ -47,5 +59,4 @@ add_custom_target(_internal_cook_game_assets ALL
        DEPENDS TheChef
 )

-
-
+destrum_cook_engine_assets(lightkeeper "${CMAKE_CURRENT_BINARY_DIR}")
--- a/lightkeeper/assets_src/plane.glb
+++ b/lightkeeper/assets_src/plane.glb
--- a/lightkeeper/src/Lightkeeper.cpp
+++ b/lightkeeper/src/Lightkeeper.cpp
@@ -11,6 +11,7 @@
 #include "destrum/Components/OrbitAndSpin.h"
 #include "destrum/ObjectModel/GameObject.h"
 #include "destrum/Util/ModelLoader.h"
+#include "destrum/Components/Animator.h"

 LightKeeper::LightKeeper(): App(), renderer(meshCache, materialCache) {
 }
@@ -52,19 +53,19 @@ void LightKeeper::customInit() {
    const float orbitRadius = 5.0f;

    for (int i = 0; i < count; ++i) {
-        auto childCube = std::make_shared<GameObject>(fmt::format("ChildCube{}", i));
-
-        auto childMeshComp = childCube->AddComponent<MeshRendererComponent>();
-        childMeshComp->SetMeshID(testMeshID);
-        childMeshComp->SetMaterialID(testMaterialID);
-
-        childCube->GetTransform().SetWorldScale(glm::vec3(0.1f));
-
-        // Add orbit + self spin
-        auto orbit = childCube->AddComponent<OrbitAndSpin>(orbitRadius, glm::vec3(0.0f));
-        orbit->Randomize(1337u + (uint32_t)i); // stable random per index
-
-        scene.Add(childCube);
+        // auto childCube = std::make_shared<GameObject>(fmt::format("ChildCube{}", i));
+        //
+        // auto childMeshComp = childCube->AddComponent<MeshRendererComponent>();
+        // childMeshComp->SetMeshID(testMeshID);
+        // childMeshComp->SetMaterialID(testMaterialID);
+        //
+        // childCube->GetTransform().SetWorldScale(glm::vec3(0.1f));
+        //
+        // // Add orbit + self spin
+        // auto orbit = childCube->AddComponent<OrbitAndSpin>(orbitRadius, glm::vec3(0.0f));
+        // orbit->Randomize(1337u + (uint32_t)i); // stable random per index
+        //
+        // scene.Add(childCube);
    }
    testCube->AddComponent<Spinner>(glm::vec3(0, 1, 0), glm::radians(10.0f)); // spin around Y, rad/sec
    //rotate 180 around X axis
@@ -94,6 +95,68 @@ void LightKeeper::customInit() {
    skyboxCubemap->CreateCubeMap();

    renderer.setSkyboxTexture(skyboxCubemap->GetCubeMapImageID());
+
+
+    const auto planeObj = std::make_shared<GameObject>("GroundPlane");
+    const auto planeMeshComp = planeObj->AddComponent<MeshRendererComponent>();
+    const auto planeModel = ModelLoader::LoadGLTF_CPUMeshes_MergedPerMesh(AssetFS::GetInstance().GetFullPath("game://plane.glb").generic_string());
+    const auto planeMeshID = meshCache.addMesh(gfxDevice, planeModel[0]);
+
+    const auto planeTextureID = gfxDevice.loadImageFromFile(AssetFS::GetInstance().GetFullPath("game://grass.png"));
+    const auto planeMaterialID = materialCache.addMaterial(gfxDevice, {
+        .baseColor = glm::vec3(1.f),
+        .textureFilteringMode = TextureFilteringMode::Nearest,
+        .diffuseTexture = planeTextureID,
+        .name = "GroundPlaneMaterial",
+    });
+    planeMeshComp->SetMeshID(planeMeshID);
+    planeMeshComp->SetMaterialID(planeMaterialID);
+    planeObj->GetTransform().SetWorldPosition(glm::vec3(0.f, -1.0f, 0.f));
+    planeObj->GetTransform().SetWorldScale(glm::vec3(10.f, 1.f, 10.f));
+    scene.Add(planeObj);
+
+
+    // At the bottom of customInit(), replace the incomplete CharObj block:
+
+    const auto CharObj = std::make_shared<GameObject>("Character");
+
+    auto charModel = ModelLoader::LoadSkinnedModel(
+        AssetFS::GetInstance().GetFullPath("engine://char.fbx").generic_string()
+    );
+
+    const auto charMeshID = meshCache.addMesh(gfxDevice, charModel.meshes[0]);
+
+    const auto charTextureID = gfxDevice.loadImageFromFile(
+        AssetFS::GetInstance().GetFullPath("engine://char.jpg"));
+    const auto charMaterialID = materialCache.addMaterial(gfxDevice, {
+        .baseColor = glm::vec3(1.f),
+        .diffuseTexture = charTextureID,
+        .name = "CharacterMaterial",
+    });
+
+    const auto charMeshComp = CharObj->AddComponent<MeshRendererComponent>();
+    charMeshComp->SetMeshID(charMeshID);
+    charMeshComp->SetMaterialID(charMaterialID);
+
+
+    const auto animator = CharObj->AddComponent<Animator>();
+    animator->setSkeleton(std::move(charModel.skeleton));
+    for (auto& clip : charModel.animations) {
+        animator->addClip(std::make_shared<SkeletalAnimation>(std::move(clip)));
+    }
+
+    for (const auto& clip : charModel.animations)
+        spdlog::info("Loaded animation: '{}' ({:.2f}s)", clip.name, clip.duration);
+
+    if (!charModel.animations.empty())
+        // animator->play(charModel.animations[0].name);
+        // animator->play("Armature|main");
+        animator->play("Armature|mixamo.com");
+    // or: animator->play("Run", 0.2f);  // 0.2s cross-fade
+
+    CharObj->GetTransform().SetWorldPosition(glm::vec3(0.f, 0.f, 0.f));
+    CharObj->GetTransform().SetWorldScale(0.01f, 0.01f, 0.01f);
+    scene.Add(CharObj);
 }

 void LightKeeper::customUpdate(float dt) {