feat: DDS 纹理加载器（BC 压缩直传 + mip 链，参考 Horizon 实现）

- Assets::Dds 解析器：中性 DdsFormat 枚举（Assets 层禁 Vulkan 头铁律），
  支持 legacy fourcc（DXT1/3/5/ATI1/2）+ DX10 头（BC1..BC7/RGBA8/BGRA8）
  + 未压缩 32bpp + 2D mip 链；数据不解码，直接上传由硬件采样解码
- Texture：按扩展名分发（.dds → InitFromDds）；渲染侧负责格式映射与
  sRGB 语义（颜色图走 SRGB 变体，BC4/5/6H 数据图保持线性）；view/sampler
  覆盖完整 mip 链；逐 mip copyBufferToImage（紧凑排布）
- Model::GetOrLoadTexture：FBX/DCC 路径反斜杠归一化（否则 Linux 全挂）
- DescriptorManager 池 1000 → 4096（SunTemple 1056 材质集打爆过；
  材质去重属材质系统任务）
- main.cpp 临时挂 SunTemple 场景验证：48 张 DDS 全部加载（2048²、12 mips），
  1056 mesh 渲染正常，validation 零警告

App/main.cpp

@@ -47,7 +47,8 @@ protected:
         EnableImGui();
 
         // 模型资产（AssetManager 负责加载/缓存/生命周期，返回句柄）
-        const auto backpack = GetAssets().LoadModel("assets/models/backpack/backpack.obj");
+        // SunTemple：FBX + DDS 压缩贴图（DdsLoader 验证场景）
+        const auto temple = GetAssets().LoadModel("assets/models/SunTemple/SunTemple.fbx");
 
         // 地面：程序化纯色平面（半边长 15，中心 y=-2）
         const auto ground = GetAssets().CreatePlane(15.0f, 140, 140, 145);
@@ -64,23 +65,13 @@ protected:
         camera.primary = true;
         camera.camera.SetPosition(glm::vec3(0.0f, 0.0f, 7.0f));
 
-        // 背包实体 × 3（悬浮在地面上方）
+        // 神庙实体 × 1（ FBX 单位较大，缩放到场景尺度）
-        auto size = 1;
-        for (int i = 0; i < size; ++i)
         {
-            for (int j = 0; j < size; j++)
+            auto e = scene.CreateEntity("SunTemple");
-            {
-                auto e = scene.CreateEntity("Backpack" + std::to_string(i));
-
             e.AddComponent<TransformComponent>()
-                 .SetPosition(glm::vec3((i - size * 0.5) * 1.5f, 0.0f, (j - size * 0.5) * 1.5f))
+             .SetScale(glm::vec3(0.05f))
-                 .SetScale(glm::vec3(0.5f))
              .Compose();
-
+            e.AddComponent<MeshRendererComponent>(MeshRendererComponent{temple});
-                e.AddComponent<MeshRendererComponent>(MeshRendererComponent{backpack});
-
-                m_Spinners.push_back(e);
-            }
         }
 
         // 灯光实体（第一个灯 = 阴影光源）

Engine/include/Engine/Assets/DdsLoader.h

@@ -0,0 +1,55 @@
+//
+// Created by atdunbg on 2026/9/23.
+//
+// DDS 文件解析（最小实现，不解码压缩块）。
+//
+// 设计：GPU 原生支持 BC 压缩格式——DDS 的压缩数据**直接上传**为
+//      BC1/BC2/BC3/BC4/BC5/BC6H/BC7 纹理由硬件采样解码，不做 CPU 解码。
+//      输出中性 DdsFormat（本层禁 Vulkan 头），vk::Format 映射与 sRGB
+//      语义由调用方（Texture，渲染侧）决定。
+//
+// 支持：
+//   - legacy fourcc：DXT1/DXT3/DXT5/ATI1/BC4U/ATI2/BC5U
+//   - DX10 扩展头：BC1..BC7、R8G8B8A8、B8G8R8A8
+//   - 未压缩 32bpp（DDPF_RGB | DDPF_ALPHAPIXELS）
+//   - 2D + mip 链（cubemap 面标志暂不支持，报错）
+//
+// 输出：Mips[r] = 第 r 级完整数据（BC 按块行紧凑排布：
+//       bytes = ceil(w/4)*ceil(h/4)*blockBytes）。
+//
+
+#ifndef HELLOVULKAN_DDSLOADER_H
+#define HELLOVULKAN_DDSLOADER_H
+
+#include <Engine/Core/Types.h>
+
+#include <string>
+#include <vector>
+
+namespace Engine::Assets::Dds {
+
+    // 中性格式（Assets 层不感知 Vulkan；映射在渲染侧 Texture）
+    enum class DdsFormat {
+        BC1, BC2, BC3, BC4, BC5, BC6H, BC7,
+        RGBA8,
+        BGRA8,   // 按字节序上传后视作 RGBA（近似，触发警告）
+    };
+
+    struct DdsData {
+        u32 Width = 0;
+        u32 Height = 0;
+        u32 MipLevels = 1;
+        DdsFormat Format = DdsFormat::RGBA8;
+        std::vector<std::vector<u8>> Mips;   // 每级完整数据（紧凑排布）
+        bool BlockCompressed = false;        // BC 格式（rowPitch 按块计算）
+    };
+
+    // 解析 DDS 文件；失败返回 false 并写 errMsg
+    bool Load(const std::string& path, DdsData& out, std::string& errMsg);
+
+    // 指定 mip 级的单行字节跨度（BC = 块行；RGBA8 = w*4）
+    u32 RowPitch(const DdsData& data, u32 mipLevel);
+
+} // namespace Engine::Assets::Dds
+
+#endif //HELLOVULKAN_DDSLOADER_H

Engine/include/Engine/ECS/Components.h

@@ -50,12 +50,12 @@ namespace Engine::ECS {
         // 分量 → world 矩阵（T * R * S）
         TransformComponent& Compose()
         {
-            glm::mat4 t = glm::translate(glm::mat4(1.0f), position);
+            const glm::mat4 t = glm::translate(glm::mat4(1.0f), position);
-            glm::mat4 r =
+            const glm::mat4 r =
                 glm::rotate(glm::mat4(1.0f), glm::radians(rotation.y), glm::vec3(0, 1, 0)) *
                 glm::rotate(glm::mat4(1.0f), glm::radians(rotation.x), glm::vec3(1, 0, 0)) *
                 glm::rotate(glm::mat4(1.0f), glm::radians(rotation.z), glm::vec3(0, 0, 1));
-            glm::mat4 s = glm::scale(glm::mat4(1.0f), scale);
+            const glm::mat4 s = glm::scale(glm::mat4(1.0f), scale);
             world = t * r * s;
             return *this;
         }

Engine/include/Engine/Vulkan/DescriptorManager.h

@@ -56,7 +56,10 @@ namespace Engine {
         Device& m_Device;
         vk::DescriptorPool m_Pool;
 
-        static constexpr u32 kMaxSets = 1000;
+        // 描述符池容量：需 ≥ 场景全部材质集合数 + 每帧集合。
+        // SunTemple（984 mesh 材质）把 1000 打爆过；材质去重（同贴图共享
+        // Material）属于材质系统任务，在那之前留足余量
+        static constexpr u32 kMaxSets = 4096;
     };
 
 } // namespace Engine

Engine/include/Engine/Vulkan/Texture.h

@@ -21,7 +21,7 @@ namespace Engine {
 
     class Texture {
     public:
-        // 从磁盘图片加载
+        // 从磁盘图片加载（按扩展名分发：.dds → DdsLoader，其它 → stbi）
         Texture(Device& device, MemoryAllocator& allocator, const std::string& path);
 
         // 程序化纹理：纯色（1x1），用于地面/调试等不依赖磁盘资产的场景
@@ -38,6 +38,8 @@ namespace Engine {
         vk::ImageView GetView() const { return m_View; }
         u32 GetWidth() const { return m_Width; }
         u32 GetHeight() const { return m_Height; }
+        vk::Format GetFormat() const { return m_Format; }
+        u32 GetMipLevels() const { return m_MipLevels; }
 
         vk::Sampler GetSampler() const { return m_Sampler; }
 
@@ -46,13 +48,19 @@ namespace Engine {
         Texture(Device& device, MemoryAllocator& allocator, ProceduralTag,
                 const std::array<u8, 4>& rgba);
 
-        // 公共路径：staging → device local image → view/sampler
+        // 公共路径：staging → device local image → view/sampler（RGBA8 单 mip）
         void InitFromPixels(const void* pixels, u32 width, u32 height);
 
+        // DDS 路径：压缩数据直传（BC 硬件采样解码）+ 文件自带 mip 链
+        void InitFromDds(const std::string& path);
+
         Device& m_Device;
         vk::Image m_Image;
         vk::ImageView m_View;
 
+        vk::Format m_Format = vk::Format::eUndefined;
+        u32 m_MipLevels = 1;
+
         vk::Sampler m_Sampler;
 
 

Engine/src/Assets/DdsLoader.cpp

@@ -0,0 +1,170 @@
+//
+// Created by atdunbg on 2026/9/23.
+//
+
+#include <Engine/Assets/DdsLoader.h>
+#include <Engine/Core/Log.h>
+
+#include <algorithm>
+#include <cstring>
+#include <fstream>
+
+namespace Engine::Assets::Dds
+{
+    namespace
+    {
+        constexpr u32 kMagic = 0x20534444;   // "DDS "
+
+        // 头部偏移（dwSpfFourCC @ 84）
+        constexpr size_t kOffsetHeight = 12;
+        constexpr size_t kOffsetWidth  = 16;
+        constexpr size_t kOffsetMips   = 28;
+        constexpr size_t kOffsetFourCC = 84;
+        constexpr size_t kHeaderSize   = 128;
+        constexpr size_t kDx10Size     = 20;
+
+        u32 ReadU32(const u8* p) { return *reinterpret_cast<const u32*>(p); }
+
+        // legacy fourcc → 中性格式
+        bool MapFourCC(u32 fourcc, DdsFormat& out)
+        {
+            switch (fourcc) {
+                case 0x31545844: out = DdsFormat::BC1; return true;  // "DXT1"
+                case 0x33545844: out = DdsFormat::BC2; return true;  // "DXT3"
+                case 0x35545844: out = DdsFormat::BC3; return true;  // "DXT5"
+                case 0x31495441: out = DdsFormat::BC4; return true;  // "ATI1"
+                case 0x55344342: out = DdsFormat::BC4; return true;  // "BC4U"
+                case 0x32495441: out = DdsFormat::BC5; return true;  // "ATI2"
+                case 0x55354342: out = DdsFormat::BC5; return true;  // "BC5U"
+                default: return false;
+            }
+        }
+
+        // DX10 DXGI 值 → 中性格式
+        bool MapDxgi(u32 dxgi, DdsFormat& out)
+        {
+            switch (dxgi) {
+                case 70: case 71: out = DdsFormat::BC1; return true;   // BC1
+                case 73: case 74: out = DdsFormat::BC2; return true;   // BC2
+                case 77: case 78: out = DdsFormat::BC3; return true;   // BC3
+                case 79: case 80: out = DdsFormat::BC4; return true;   // BC4
+                case 82: case 83: out = DdsFormat::BC5; return true;   // BC5
+                case 94: case 95: out = DdsFormat::BC6H; return true;  // BC6H
+                case 97: case 98: out = DdsFormat::BC7; return true;   // BC7
+                case 28: case 29: out = DdsFormat::RGBA8; return true; // R8G8B8A8
+                case 87: case 91: out = DdsFormat::BGRA8; return true; // B8G8R8A8（近似）
+                default: return false;
+            }
+        }
+
+        u32 BytesPerBlock(DdsFormat format)
+        {
+            switch (format) {
+                case DdsFormat::BC1:
+                case DdsFormat::BC4:
+                    return 8;
+                case DdsFormat::BC2:
+                case DdsFormat::BC3:
+                case DdsFormat::BC5:
+                case DdsFormat::BC6H:
+                case DdsFormat::BC7:
+                    return 16;
+                default:
+                    return 4;   // RGBA8/BGRA8 系
+            }
+        }
+
+        bool IsBlockCompressed(DdsFormat format)
+        {
+            return format != DdsFormat::RGBA8 && format != DdsFormat::BGRA8;
+        }
+    }
+
+    bool Load(const std::string& path, DdsData& out, std::string& errMsg)
+    {
+        std::ifstream f(path, std::ios::binary);
+        if (!f) {
+            errMsg = "无法打开文件";
+            return false;
+        }
+
+        u8 header[kHeaderSize];
+        f.read(reinterpret_cast<char*>(header), kHeaderSize);
+        if (!f || ReadU32(header) != kMagic) {
+            errMsg = "DDS 魔数不符";
+            return false;
+        }
+
+        out.Width  = ReadU32(header + kOffsetWidth);
+        out.Height = ReadU32(header + kOffsetHeight);
+        const u32 mips = ReadU32(header + kOffsetMips);
+        out.MipLevels = mips ? mips : 1;
+
+        const u32 fourcc = ReadU32(header + kOffsetFourCC);
+        size_t dataOffset = kHeaderSize;
+
+        if (fourcc == 0x30313044) {   // "DX10"
+            u8 dx10[kDx10Size];
+            f.read(reinterpret_cast<char*>(dx10), kDx10Size);
+            if (!f) {
+                errMsg = "DX10 头不完整";
+                return false;
+            }
+            const u32 dxgi = ReadU32(dx10);
+            if (!MapDxgi(dxgi, out.Format)) {
+                errMsg = "DX10 DXGI 格式不支持";
+                return false;
+            }
+            dataOffset += kDx10Size;
+        } else if (!MapFourCC(fourcc, out.Format)) {
+            // 未压缩路径：仅支持 32bpp RGBA（DDPF_ALPHAPIXELS(0x1) | DDPF_RGB(0x40)）
+            const u32 ddspfFlags = ReadU32(header + 80);
+            if (ddspfFlags & 0x41) {
+                out.Format = DdsFormat::BGRA8;   // 常见 32bpp DDS 是 B8G8R8A8 字节序
+                ENGINE_WARN("DDS: 未压缩 32bpp 按 BGRA8 上传 '{}'", path);
+            } else {
+                errMsg = "fourcc 不支持";
+                return false;
+            }
+        }
+
+        out.BlockCompressed = IsBlockCompressed(out.Format);
+        f.seekg(dataOffset);
+
+        // 数据量校验 + 逐级读取（紧凑排布）
+        const u32 bpb  = BytesPerBlock(out.Format);
+        const u32 cols = out.BlockCompressed ? 4 : 1;
+        out.Mips.resize(out.MipLevels);
+
+        u32 w = out.Width, h = out.Height;
+        for (u32 mip = 0; mip < out.MipLevels; ++mip) {
+            const u32 blocksW = std::max(1u, (w + cols - 1) / cols);
+            const u32 blocksH = std::max(1u, (h + cols - 1) / cols);
+            const size_t bytes = size_t(blocksW) * blocksH * bpb;
+            out.Mips[mip].resize(bytes);
+            f.read(reinterpret_cast<char*>(out.Mips[mip].data()),
+                   std::streamsize(bytes));
+            if (!f) {
+                errMsg = "mip 数据不完整";
+                return false;
+            }
+            w = std::max(1u, w / 2);
+            h = std::max(1u, h / 2);
+        }
+        return true;
+    }
+
+    u32 RowPitch(const DdsData& data, u32 mipLevel)
+    {
+        u32 w = data.Width;
+        for (u32 m = 0; m < mipLevel; ++m) {
+            w = std::max(1u, w / 2);
+        }
+        const u32 bpb = BytesPerBlock(data.Format);
+        if (data.BlockCompressed) {
+            return std::max(1u, (w + 3) / 4) * bpb;
+        }
+        return w * bpb;
+    }
+
+} // namespace Engine::Assets::Dds

Engine/src/Renderer/Model.cpp

@@ -15,6 +15,7 @@
 #include <assimp/postprocess.h>
 
 #include <filesystem>
+#include <algorithm>
 
 namespace Engine {
 
@@ -110,8 +111,12 @@ void Model::ProcessNode(aiNode* node, const aiScene* scene)
     }
 }
 
-Texture* Model::GetOrLoadTexture(const std::string& path)
+Texture* Model::GetOrLoadTexture(const std::string& rawPath)
 {
+    // FBX/DCC 导出的路径常带反斜杠（Windows 语义），统一成正斜杠
+    std::string path = rawPath;
+    std::replace(path.begin(), path.end(), '\\', '/');
+
     auto it = m_TextureCache.find(path);
     if (it != m_TextureCache.end()) {
         return it->second.get();

Engine/src/Vulkan/Texture.cpp

@@ -5,10 +5,13 @@
 #include <Engine/Vulkan/Texture.h>
 #include <Engine/Vulkan/Device.h>
 #include <Engine/Vulkan/MemoryAllocator.h>
+#include <Engine/Assets/DdsLoader.h>
 #include <Engine/Core/Log.h>
 
 #include <stb/stb_image.h>
 
+#include <algorithm>
+#include <cctype>
 #include <cstring>
 
 #include "Engine/Core/Stats.h"
@@ -18,12 +21,21 @@ namespace Engine
     Texture::Texture(Device& device, MemoryAllocator& allocator, const std::string& path)
         : m_Device(device) , m_Allocator(allocator)
     {
+        // 按扩展名分发：DDS（BC 压缩直传 + mip 链）vs stbi（无压缩单 mip）
+        const std::string ext = path.size() >= 4
+            ? path.substr(path.size() - 4) : "";
+        std::string lower;
+        for (char c : ext) lower += static_cast<char>(std::tolower(c));
+
+        if (lower == ".dds") {
+            InitFromDds(path);
+        } else {
             // 1. CPU 加载图片
             int width = 0, height = 0, channels = 0;
-        stbi_uc* pixels = stbi_load(path.c_str(), &width, &height, &channels, STBI_rgb_alpha);
+            stbi_uc* pixels = stbi_load(path.c_str(), &width, &height, &channels,
+                                        STBI_rgb_alpha);
 
-        if (!pixels)
+            if (!pixels) {
-        {
                 ENGINE_ERROR("Failed to load texture: {}", path);
                 return;
             }
@@ -34,6 +46,7 @@ namespace Engine
 
             ENGINE_INFO("Texture loaded: {} ({}x{})", path, m_Width, m_Height);
         }
+    }
 
     Texture::Texture(Device& device, MemoryAllocator& allocator, ProceduralTag,
                      const std::array<u8, 4>& rgba)
@@ -53,6 +66,8 @@ namespace Engine
     {
         m_Width = width;
         m_Height = height;
+        m_Format = vk::Format::eR8G8B8A8Srgb;
+        m_MipLevels = 1;
         vk::DeviceSize imageSize = static_cast<vk::DeviceSize>(width) *
             static_cast<vk::DeviceSize>(height) * 4; // RGBA8
 
@@ -87,10 +102,10 @@ namespace Engine
         vk::ImageViewCreateInfo viewInfo;
         viewInfo.setImage(m_Image)
                 .setViewType(vk::ImageViewType::e2D)
-                .setFormat(vk::Format::eR8G8B8A8Srgb)
+                .setFormat(m_Format)
                 .setSubresourceRange({
                     vk::ImageAspectFlagBits::eColor,
-                    0, 1, 0, 1
+                    0, m_MipLevels, 0, 1
                 });
 
         m_View = m_Device.GetDevice().createImageView(viewInfo);
@@ -107,7 +122,9 @@ namespace Engine
                    .setBorderColor(vk::BorderColor::eIntOpaqueBlack)
                    .setUnnormalizedCoordinates(VK_FALSE)
                    .setCompareEnable(VK_FALSE)
-                   .setMipmapMode(vk::SamplerMipmapMode::eLinear);
+                   .setMipmapMode(vk::SamplerMipmapMode::eLinear)
+                   .setMinLod(0.0f)
+                   .setMaxLod(static_cast<f32>(m_MipLevels));
 
         m_Sampler = m_Device.GetDevice().createSampler(samplerInfo);
 
@@ -197,6 +214,186 @@ namespace Engine
         STATS.AddTexture(m_TextureBytes);
     }
 
+    // DDS → Vulkan 格式映射（渲染侧职责：sRGB 语义按颜色贴图默认开启；
+    // BC4/BC5/BC6H 是数据图，保持线性）
+    vk::Format ToVkFormat(Assets::Dds::DdsFormat f)
+    {
+        using DF = Assets::Dds::DdsFormat;
+        switch (f) {
+            case DF::BC1:   return vk::Format::eBc1RgbaSrgbBlock;
+            case DF::BC2:   return vk::Format::eBc2SrgbBlock;
+            case DF::BC3:   return vk::Format::eBc3SrgbBlock;
+            case DF::BC4:   return vk::Format::eBc4UnormBlock;
+            case DF::BC5:   return vk::Format::eBc5UnormBlock;
+            case DF::BC6H:  return vk::Format::eBc6HUfloatBlock;
+            case DF::BC7:   return vk::Format::eBc7SrgbBlock;
+            case DF::RGBA8: return vk::Format::eR8G8B8A8Srgb;
+            case DF::BGRA8: return vk::Format::eB8G8R8A8Srgb;
+        }
+        return vk::Format::eUndefined;
+    }
+
+    u32 MipDim(u32 v, u32 level)
+    {
+        return std::max(1u, v >> level);
+    }
+
+    void Texture::InitFromDds(const std::string& path)
+    {
+        Assets::Dds::DdsData data;
+        std::string err;
+        if (!Assets::Dds::Load(path, data, err)) {
+            ENGINE_ERROR("DDS: {} ({})", err, path);
+            return;
+        }
+
+        m_Width = data.Width;
+        m_Height = data.Height;
+        m_MipLevels = data.MipLevels;
+        m_Format = ToVkFormat(data.Format);
+
+        ENGINE_INFO("DDS loaded: {} ({}x{}, {} mips, {})", path,
+                    m_Width, m_Height, m_MipLevels, vk::to_string(m_Format));
+
+        // 1. staging：各 mip 紧凑排布，记偏移
+        vk::DeviceSize total = 0;
+        for (const auto& mip : data.Mips) {
+            total += mip.size();
+        }
+
+        auto stagingAlloc = m_Allocator.CreateBuffer(
+            total,
+            vk::BufferUsageFlagBits::eTransferSrc,
+            vk::MemoryPropertyFlagBits::eHostVisible |
+            vk::MemoryPropertyFlagBits::eHostCoherent);
+
+        std::vector<vk::DeviceSize> offsets(m_MipLevels);
+        {
+            vk::DeviceSize off = 0;
+            for (u32 mip = 0; mip < m_MipLevels; ++mip) {
+                offsets[mip] = off;
+                std::memcpy(stagingAlloc.MappedPtr + off,
+                            data.Mips[mip].data(), data.Mips[mip].size());
+                off += data.Mips[mip].size();
+            }
+        }
+
+        // 2. device local image（带完整 mip 链）
+        vk::ImageCreateInfo imageInfo;
+        imageInfo.setImageType(vk::ImageType::e2D)
+                 .setFormat(m_Format)
+                 .setExtent({m_Width, m_Height, 1})
+                 .setMipLevels(m_MipLevels)
+                 .setArrayLayers(1)
+                 .setSamples(vk::SampleCountFlagBits::e1)
+                 .setTiling(vk::ImageTiling::eOptimal)
+                 .setUsage(vk::ImageUsageFlagBits::eTransferDst |
+                           vk::ImageUsageFlagBits::eSampled)
+                 .setSharingMode(vk::SharingMode::eExclusive)
+                 .setInitialLayout(vk::ImageLayout::eUndefined);
+
+        m_ImageAllocation = m_Allocator.CreateImage(imageInfo,
+            vk::MemoryPropertyFlagBits::eDeviceLocal);
+        m_Image = m_ImageAllocation.Image;
+
+        // 3. view / sampler（覆盖整个 mip 链）
+        m_View = m_Device.GetDevice().createImageView(
+            vk::ImageViewCreateInfo{}
+                .setImage(m_Image)
+                .setViewType(vk::ImageViewType::e2D)
+                .setFormat(m_Format)
+                .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                      0, m_MipLevels, 0, 1}));
+
+        m_Sampler = m_Device.GetDevice().createSampler(
+            vk::SamplerCreateInfo{}
+                .setMagFilter(vk::Filter::eLinear)
+                .setMinFilter(vk::Filter::eLinear)
+                .setMipmapMode(vk::SamplerMipmapMode::eLinear)
+                .setAddressModeU(vk::SamplerAddressMode::eRepeat)
+                .setAddressModeV(vk::SamplerAddressMode::eRepeat)
+                .setAddressModeW(vk::SamplerAddressMode::eRepeat)
+                .setAnisotropyEnable(VK_FALSE)
+                .setBorderColor(vk::BorderColor::eIntOpaqueBlack)
+                .setUnnormalizedCoordinates(VK_FALSE)
+                .setCompareEnable(VK_FALSE)
+                .setMinLod(0.0f)
+                .setMaxLod(static_cast<f32>(m_MipLevels)));
+
+        // 4. 一次性命令缓冲：整图 → TransferDst，逐 mip 拷贝 → ShaderRead
+        vk::CommandPoolCreateInfo poolInfo;
+        poolInfo.setQueueFamilyIndex(m_Device.GetGraphicsFamily())
+                .setFlags(vk::CommandPoolCreateFlagBits::eTransient);
+        vk::CommandPool pool = m_Device.GetDevice().createCommandPool(poolInfo);
+
+        vk::CommandBufferAllocateInfo allocInfo;
+        allocInfo.setCommandPool(pool)
+                 .setLevel(vk::CommandBufferLevel::ePrimary)
+                 .setCommandBufferCount(1);
+        vk::CommandBuffer cmd =
+            m_Device.GetDevice().allocateCommandBuffers(allocInfo)[0];
+
+        cmd.begin(vk::CommandBufferBeginInfo{
+            vk::CommandBufferUsageFlagBits::eOneTimeSubmit});
+
+        vk::ImageMemoryBarrier2 toTransfer;
+        toTransfer.setSrcStageMask(vk::PipelineStageFlagBits2::eNone)
+                  .setSrcAccessMask(vk::AccessFlagBits2::eNone)
+                  .setDstStageMask(vk::PipelineStageFlagBits2::eTransfer)
+                  .setDstAccessMask(vk::AccessFlagBits2::eTransferWrite)
+                  .setOldLayout(vk::ImageLayout::eUndefined)
+                  .setNewLayout(vk::ImageLayout::eTransferDstOptimal)
+                  .setImage(m_Image)
+                  .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                        0, m_MipLevels, 0, 1});
+        cmd.pipelineBarrier2(vk::DependencyInfo{}.setImageMemoryBarriers(toTransfer));
+
+        // BC 数据按块紧凑排布 → bufferRowLength/Height 留 0（紧凑）
+        for (u32 mip = 0; mip < m_MipLevels; ++mip) {
+            vk::BufferImageCopy region;
+            region.setBufferOffset(offsets[mip])
+                  .setBufferRowLength(0)
+                  .setBufferImageHeight(0)
+                  .setImageSubresource({vk::ImageAspectFlagBits::eColor,
+                                        mip, 0, 1})
+                  .setImageOffset({0, 0, 0})
+                  .setImageExtent({MipDim(m_Width, mip),
+                                   MipDim(m_Height, mip), 1});
+            cmd.copyBufferToImage(stagingAlloc.Buffer, m_Image,
+                                  vk::ImageLayout::eTransferDstOptimal, region);
+        }
+
+        vk::ImageMemoryBarrier2 toShader;
+        toShader.setSrcStageMask(vk::PipelineStageFlagBits2::eTransfer)
+                .setSrcAccessMask(vk::AccessFlagBits2::eTransferWrite)
+                .setDstStageMask(vk::PipelineStageFlagBits2::eFragmentShader)
+                .setDstAccessMask(vk::AccessFlagBits2::eShaderRead)
+                .setOldLayout(vk::ImageLayout::eTransferDstOptimal)
+                .setNewLayout(vk::ImageLayout::eShaderReadOnlyOptimal)
+                .setImage(m_Image)
+                .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                      0, m_MipLevels, 0, 1});
+        cmd.pipelineBarrier2(vk::DependencyInfo{}.setImageMemoryBarriers(toShader));
+
+        cmd.end();
+
+        // 5. 提交并等待，清理 staging
+        vk::Fence fence = m_Device.GetDevice().createFence({});
+        m_Device.GetGraphicsQueue().submit(
+            vk::SubmitInfo{}.setCommandBuffers(cmd), fence);
+        if (m_Device.GetDevice().waitForFences(fence, VK_TRUE, UINT64_MAX) !=
+            vk::Result::eSuccess) {
+            ENGINE_ERROR("DDS upload: waitForFences failed");
+        }
+
+        m_Device.GetDevice().destroyFence(fence);
+        m_Device.GetDevice().destroyCommandPool(pool);
+        m_Allocator.Destroy(stagingAlloc);
+
+        m_TextureBytes = total;
+        STATS.AddTexture(m_TextureBytes);
+    }
+
     Texture::~Texture()
     {
         const auto& vkDevice = m_Device.GetDevice();