feat: IBL 阶段②——漫反射辐照度卷积，PBR ambient 接入环境光

- Cubemap 构造期同步 CPU 半球 Lambert 积分（64×32 网格 ≈2048 采样/像素）
  生成 32² irradiance cube（漫反射是超低频信号，分辨率无需高）
- shader：kD 用粗糙度感知 FresnelSchlickRoughness + (1-metallic)，
  N 采样 irradianceMap（Set2 binding 3）替换常数 ambient；
  ambientColor 保留为强度/色调乘子
- CreateNeutral 兜底：无 HDR 资产时 1×1 常数 cube，描述符恒有效
- FaceDir 改为返回单位方向（equirect 查表要求，修掉未归一化的畸变）

Engine/include/Engine/Renderer/Cubemap.h

@@ -1,11 +1,19 @@
 //
 // Created by atdunbg on 2026/9/23.
 //
-// Cubemap：环境立方体贴图。
+// Cubemap：环境立方体贴图 + 漫反射辐照度图（IBL 阶段②）。
 //
 // 从等距柱状 HDR（.hdr，polyhaven 等）生成：CPU 重采样到 6 面
 //（一次性加载成本，亚秒级），转 fp16 后整块上传，硬件 Cube 采样。
-// 后续 IBL 的 irradiance/prefilter 卷积也基于本类的数据形态。
+// 同一次加载内完成 irradiance 卷积（半球积分，源数据在手零额外 IO）。
+//
+// 双图契约：
+//   Get()/GetView()          —— 环境 cube（Skybox 背景）
+//   GetIrradianceView()      —— 漫反射辐照度 cube（PBR ambient）
+// 镜面 IBL（pre-filter + BRDF LUT）后续接入，形态一致。
+//
+// 后续演进：CPU 卷积 → compute pass（动态环境零卡顿）；本类的
+// 方向表/上传路径届时整体迁移，接口不变。
 //
 
 #ifndef HELLOVULKAN_CUBEMAP_H
@@ -13,6 +21,7 @@
 
 #include <Engine/Core/Types.h>
 #include <Engine/Vulkan/MemoryAllocator.h>
+#include <glm/glm.hpp>
 #include <vulkan/vulkan.hpp>
 
 namespace Engine {
@@ -21,36 +30,55 @@ namespace Engine {
 
     class Cubemap {
     public:
-        // 从等距柱状 HDR 生成（CPU 重采样到 6 面，faceSize 为每面边长）
+        // 从等距柱状 HDR 生成（CPU 重采样 6 面 + irradiance 卷积）
         static Scope<Cubemap> CreateFromHdrEquirect(Device& device,
                                                     MemoryAllocator& allocator,
                                                     const std::string& path,
                                                     u32 faceSize);
 
+        // 中性兜底（无 HDR 资产时）：env/irradiance 均为常数色。
+        // 保证引用侧描述符恒有效（validation），渲染退化为常数环境
+        static Scope<Cubemap> CreateNeutral(Device& device,
+                                            MemoryAllocator& allocator,
+                                            const glm::vec3& rgb);
+
         ~Cubemap();
 
         Cubemap(const Cubemap&) = delete;
         Cubemap& operator=(const Cubemap&) = delete;
 
-        vk::Image Get() const { return m_Image; }
-        vk::ImageView GetView() const { return m_View; }
+        // ---- 环境 cube（Skybox 背景）----
+        vk::Image Get() const { return m_EnvImage; }
+        vk::ImageView GetView() const { return m_EnvView; }
+
+        // ---- 漫反射辐照度 cube（PBR ambient）----
+        vk::ImageView GetIrradianceView() const { return m_IrrView; }
+
         vk::Sampler GetSampler() const { return m_Sampler; }
         u32 GetFaceSize() const { return m_FaceSize; }
 
     private:
         Cubemap(Device& device, MemoryAllocator& allocator,
-                vk::Image image, vk::ImageView view, vk::Sampler sampler,
-                MemoryAllocator::Allocation allocation, u32 faceSize)
+                vk::Image envImage, vk::ImageView envView,
+                vk::Image irrImage, vk::ImageView irrView, vk::Sampler sampler,
+                MemoryAllocator::Allocation envAlloc,
+                MemoryAllocator::Allocation irrAlloc, u32 faceSize)
             : m_Device(device), m_Allocator(allocator),
-              m_Image(image), m_View(view), m_Sampler(sampler),
-              m_Allocation(allocation), m_FaceSize(faceSize) {}
+              m_EnvImage(envImage), m_EnvView(envView),
+              m_IrrImage(irrImage), m_IrrView(irrView),
+              m_Sampler(sampler), m_EnvAllocation(envAlloc),
+              m_IrrAllocation(irrAlloc), m_FaceSize(faceSize) {}
 
         Device& m_Device;
         MemoryAllocator& m_Allocator;
-        vk::Image m_Image;
-        vk::ImageView m_View;
+
+        vk::Image m_EnvImage;
+        vk::ImageView m_EnvView;
+        vk::Image m_IrrImage;
+        vk::ImageView m_IrrView;
         vk::Sampler m_Sampler;
-        MemoryAllocator::Allocation m_Allocation;
+        MemoryAllocator::Allocation m_EnvAllocation;
+        MemoryAllocator::Allocation m_IrrAllocation;
         u32 m_FaceSize;
     };
 

Engine/include/Engine/Renderer/Passes/SceneGpu.h

@@ -24,6 +24,7 @@ namespace Engine {
     class DescriptorManager;
     class ShaderManager;
     class UniformBuffer;
+    class Cubemap;
 
     namespace Passes {
 
@@ -40,8 +41,8 @@ namespace Engine {
 
             // ---- set 布局（shader 契约，改绑定必须同步 shaders/）----
             vk::DescriptorSetLayout uboSetLayout;      // Set0 = {UBO, ObjectBuffer, PrevObjectBuffer}
-            vk::DescriptorSetLayout textureSetLayout;  // Set1 = {albedo}
-            vk::DescriptorSetLayout shadowSetLayout;   // Set2 = {级联 0/1/2 阴影图}
+            vk::DescriptorSetLayout textureSetLayout;  // Set1 = {材质 4 槽 + 参数 UBO}
+            vk::DescriptorSetLayout shadowSetLayout;   // Set2 = {级联 0/1/2 阴影图, irradiance cube}
             vk::PipelineLayout forwardLayout;          // 3 sets，无 push（GPUDriven 经 SSBO）
             vk::PipelineLayout shadowLayout;           // Set0，push = lightVP 64B
 
@@ -60,6 +61,9 @@ namespace Engine {
             void BindShadowMaps(RG::RenderGraph& rg,
                                 const std::array<RG::TextureHandle, kMaxCascades>& maps);
 
+            // 环境辐照度（Set2 binding 3）：view 全程稳定，绑一次即可
+            void BindEnvironment(const Cubemap& env);
+
         private:
             Device& m_Device;
             DescriptorManager& m_Descriptors;

Engine/src/Renderer/Cubemap.cpp

@@ -1,11 +1,14 @@
 //
 // Created by atdunbg on 2026/9/23.
 //
-// Cubemap 实现：HDR 等距柱状 → CPU 重采样 6 面 → fp16 → cube image。
+// Cubemap 实现：HDR 等距柱状 → CPU 重采样 6 面 + irradiance 卷积 → fp16 cube。
 //
-// 一次性加载成本（亚秒级 CPU 重采样），换取运行时零转换开销；
+// 一次性加载成本（亚秒级 CPU），换取运行时零转换开销；
 // 硬件 Cube 采样（filtering/边缘接缝由 driver 处理）。
-// 后续 IBL 阶段的 irradiance/prefilter 卷积复用 FaceDir 方向表。
+//
+// irradiance 卷积（LearnOpenGL split-sum 的离线端）：对每个法线方向
+// 做半球 Lambert 积分 E = ∫ L·cosθ·sinθ dθ dφ，切线基由 up×N 构造。
+// 32² 面足够——漫反射辐照度本来就是超低频信号。
 //
 
 #include <Engine/Renderer/Cubemap.h>
@@ -29,22 +32,27 @@ namespace Engine
     namespace
     {
         constexpr u32 kFaceCount = 6;
+        constexpr u32 kIrradianceFaceSize = 32;   // 漫反射辐照度面边长（低频）
+        constexpr u32 kPhiSteps = 64;             // 方位角采样数
+        constexpr u32 kThetaSteps = 32;           // 极角采样数
 
         // GL/Vulkan 规范的 cube 面方向表（与 GL spec 立方体选择表互逆）。
-        // (u,v)∈[0,1]（u 向右 v 向下）→ s,t∈[-1,1] → 单位化前方向。
+        // (u,v)∈[0,1]（u 向右 v 向下）→ s,t∈[-1,1] → 单位方向。
         // Vulkan 面序与 GL 一致：+X -X +Y -Y +Z -Z
         glm::vec3 FaceDir(u32 face, f32 u, f32 v)
         {
             const f32 s = u * 2.0f - 1.0f;
             const f32 t = v * 2.0f - 1.0f;
+            glm::vec3 dir;
             switch (face) {
-                case 0:  return { 1.0f, -t, -s };   // +X（面片顶行 = +Y）
-                case 1:  return {-1.0f, -t,  s };   // -X
-                case 2:  return { s, 1.0f,  t };    // +Y
-                case 3:  return { s,-1.0f, -t };    // -Y
-                case 4:  return { s, -t, 1.0f };    // +Z
-                default: return {-s, -t,-1.0f };    // -Z
+                case 0:  dir = { 1.0f, -t, -s }; break;   // +X（面片顶行 = +Y）
+                case 1:  dir = {-1.0f, -t,  s }; break;   // -X
+                case 2:  dir = { s, 1.0f,  t };  break;   // +Y
+                case 3:  dir = { s,-1.0f, -t };  break;   // -Y
+                case 4:  dir = { s, -t, 1.0f };  break;   // +Z
+                default: dir = {-s, -t,-1.0f };  break;   // -Z
             }
+            return glm::normalize(dir);   // equirect 查表要求单位方向
         }
 
         // 等距柱状双线性采样（x 环绕跨经度接缝，y 钳制到极点）。
@@ -76,6 +84,39 @@ namespace Engine
             return glm::mix(glm::mix(c00, c10, tx), glm::mix(c01, c11, tx), ty);
         }
 
+        // 漫反射辐照度半球积分（Lambert）：E = ∫ L·cosθ·sinθ dθ dφ。
+        // 切线基 up×N 构造；φSteps×θSteps 均匀网格 ≈ 2048 采样/像素，
+        // 对 32² 低频图绰绰有余
+        glm::vec3 ConvolveIrradiance(const f32* img, u32 w, u32 h,
+                                     const glm::vec3& N)
+        {
+            const glm::vec3 up = glm::abs(N.y) > 0.999f
+                ? glm::vec3(0.0f, 0.0f, 1.0f) : glm::vec3(0.0f, 1.0f, 0.0f);
+            const glm::vec3 T = glm::normalize(glm::cross(up, N));
+            const glm::vec3 B = glm::cross(N, T);
+
+            const f32 dPhi = 2.0f * glm::pi<f32>() / static_cast<f32>(kPhiSteps);
+            const f32 dTheta = 0.5f * glm::pi<f32>() / static_cast<f32>(kThetaSteps);
+
+            glm::vec3 irradiance(0.0f);
+            for (u32 p = 0; p < kPhiSteps; ++p) {
+                const f32 phi = (static_cast<f32>(p) + 0.5f) * dPhi;
+                for (u32 t = 0; t < kThetaSteps; ++t) {
+                    const f32 theta = (static_cast<f32>(t) + 0.5f) * dTheta;
+                    const f32 sinT = std::sin(theta);
+                    const f32 cosT = std::cos(theta);
+                    // 切线空间采样方向 → 世界
+                    const glm::vec3 sampleDir =
+                        T * (sinT * std::cos(phi)) +
+                        B * (sinT * std::sin(phi)) +
+                        N * cosT;
+                    irradiance += SampleEquirect(img, w, h, sampleDir) *
+                                  (cosT * sinT);
+                }
+            }
+            return irradiance * (dPhi * dTheta);
+        }
+
         // IEEE754 f32 → fp16（round-to-nearest-even；溢出饱和为 Inf）。
         // HDR 环境值域内常规路径足够，次正规/NaN 路径只为数学完备
         u16 FloatToHalf(f32 value)
@@ -106,6 +147,125 @@ namespace Engine
             // 舍入进位到 0x400 恰好是最小正规 half 的位型，无需特判
             return static_cast<u16>(sign | std::min(m, 0x400u));
         }
+
+        // 上传 6 面 fp16 数据为 cube image（UNDEFINED → TransferDst →
+        // ShaderRead；一次性命令缓冲）。view 由调用方创建
+        MemoryAllocator::Allocation UploadCube(Device& device,
+                                               MemoryAllocator& allocator,
+                                               const std::vector<u16>& faces,
+                                               u32 faceSize, vk::Image& outImage)
+        {
+            const vk::DeviceSize faceBytes =
+                static_cast<vk::DeviceSize>(faceSize) * faceSize * 8;
+            const vk::DeviceSize total = faceBytes * kFaceCount;
+
+            auto stagingAlloc = allocator.CreateBuffer(
+                total,
+                vk::BufferUsageFlagBits::eTransferSrc,
+                vk::MemoryPropertyFlagBits::eHostVisible |
+                vk::MemoryPropertyFlagBits::eHostCoherent);
+            std::memcpy(stagingAlloc.MappedPtr, faces.data(),
+                        static_cast<size_t>(total));
+
+            vk::ImageCreateInfo imageInfo;
+            imageInfo.setImageType(vk::ImageType::e2D)
+                     .setFormat(vk::Format::eR16G16B16A16Sfloat)
+                     .setExtent({faceSize, faceSize, 1})
+                     .setMipLevels(1)
+                     .setArrayLayers(kFaceCount)
+                     .setSamples(vk::SampleCountFlagBits::e1)
+                     .setTiling(vk::ImageTiling::eOptimal)
+                     .setFlags(vk::ImageCreateFlagBits::eCubeCompatible)
+                     .setUsage(vk::ImageUsageFlagBits::eTransferDst |
+                               vk::ImageUsageFlagBits::eSampled)
+                     .setSharingMode(vk::SharingMode::eExclusive)
+                     .setInitialLayout(vk::ImageLayout::eUndefined);
+
+            auto imageAlloc = allocator.CreateImage(imageInfo,
+                vk::MemoryPropertyFlagBits::eDeviceLocal);
+            outImage = imageAlloc.Image;
+
+            vk::CommandPoolCreateInfo poolInfo;
+            poolInfo.setQueueFamilyIndex(device.GetGraphicsFamily())
+                    .setFlags(vk::CommandPoolCreateFlagBits::eTransient);
+            vk::CommandPool pool = device.GetDevice().createCommandPool(poolInfo);
+
+            vk::CommandBuffer cmd = device.GetDevice().allocateCommandBuffers(
+                vk::CommandBufferAllocateInfo{}
+                    .setCommandPool(pool)
+                    .setLevel(vk::CommandBufferLevel::ePrimary)
+                    .setCommandBufferCount(1))[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(outImage)
+                      .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                            0, 1, 0, kFaceCount});
+            cmd.pipelineBarrier2(
+                vk::DependencyInfo{}.setImageMemoryBarriers(toTransfer));
+
+            for (u32 f = 0; f < kFaceCount; ++f) {
+                vk::BufferImageCopy region;
+                region.setBufferOffset(faceBytes * f)
+                      .setBufferRowLength(0)
+                      .setBufferImageHeight(0)
+                      .setImageSubresource({vk::ImageAspectFlagBits::eColor,
+                                            0, f, 1})
+                      .setImageOffset({0, 0, 0})
+                      .setImageExtent({faceSize, faceSize, 1});
+                cmd.copyBufferToImage(stagingAlloc.Buffer, outImage,
+                                      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(outImage)
+                    .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                          0, 1, 0, kFaceCount});
+            cmd.pipelineBarrier2(
+                vk::DependencyInfo{}.setImageMemoryBarriers(toShader));
+
+            cmd.end();
+
+            vk::Fence fence = device.GetDevice().createFence({});
+            device.GetGraphicsQueue().submit(
+                vk::SubmitInfo{}.setCommandBuffers(cmd), fence);
+            if (device.GetDevice().waitForFences(fence, VK_TRUE, UINT64_MAX) !=
+                vk::Result::eSuccess) {
+                ENGINE_ERROR("Cubemap upload: waitForFences failed");
+            }
+
+            device.GetDevice().destroyFence(fence);
+            device.GetDevice().destroyCommandPool(pool);
+            allocator.Destroy(stagingAlloc);
+
+            return imageAlloc;
+        }
+
+        vk::ImageView CreateCubeView(Device& device, vk::Image image)
+        {
+            return device.GetDevice().createImageView(
+                vk::ImageViewCreateInfo{}
+                    .setImage(image)
+                    .setViewType(vk::ImageViewType::eCube)
+                    .setFormat(vk::Format::eR16G16B16A16Sfloat)
+                    .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                          0, 1, 0, kFaceCount}));
+        }
     }
 
     Scope<Cubemap> Cubemap::CreateFromHdrEquirect(Device& device,
@@ -123,11 +283,11 @@ namespace Engine
         const u32 srcW = static_cast<u32>(w);
         const u32 srcH = static_cast<u32>(h);
 
-        // ---- 2. 重采样 6 面 + fp16（RGBA16F = 每像素 8B）----
+        // ---- 2. 重采样 6 面 + irradiance 卷积 + fp16 ----
         const u32 facePixels = faceSize * faceSize;
-        std::vector<u16> faces(kFaceCount * facePixels * 4);
+        std::vector<u16> envFaces(kFaceCount * facePixels * 4);
         for (u32 f = 0; f < kFaceCount; ++f) {
-            u16* dst = faces.data() + f * facePixels * 4;
+            u16* dst = envFaces.data() + f * facePixels * 4;
             for (u32 y = 0; y < faceSize; ++y) {
                 for (u32 x = 0; x < faceSize; ++x) {
                     const f32 u = (static_cast<f32>(x) + 0.5f) / static_cast<f32>(faceSize);
@@ -142,50 +302,37 @@ namespace Engine
                 }
             }
         }
-        stbi_image_free(pixels);
 
-        // ---- 3. staging（HOST_VISIBLE 持久映射，逐面偏移拷入）----
-        const vk::DeviceSize faceBytes =
-            static_cast<vk::DeviceSize>(facePixels) * 8;
-        const vk::DeviceSize total = faceBytes * kFaceCount;
+        const u32 irrPixels = kIrradianceFaceSize * kIrradianceFaceSize;
+        std::vector<u16> irrFaces(kFaceCount * irrPixels * 4);
+        for (u32 f = 0; f < kFaceCount; ++f) {
+            u16* dst = irrFaces.data() + f * irrPixels * 4;
+            for (u32 y = 0; y < kIrradianceFaceSize; ++y) {
+                for (u32 x = 0; x < kIrradianceFaceSize; ++x) {
+                    const f32 u = (static_cast<f32>(x) + 0.5f) / static_cast<f32>(kIrradianceFaceSize);
+                    const f32 v = (static_cast<f32>(y) + 0.5f) / static_cast<f32>(kIrradianceFaceSize);
+                    const glm::vec3 rgb = ConvolveIrradiance(
+                        pixels, srcW, srcH, FaceDir(f, u, v));
+                    u16* px = dst + (static_cast<size_t>(y) * kIrradianceFaceSize + x) * 4;
+                    px[0] = FloatToHalf(rgb.r);
+                    px[1] = FloatToHalf(rgb.g);
+                    px[2] = FloatToHalf(rgb.b);
+                    px[3] = 0x3C00;
+                }
+            }
+        }
+        stbi_image_free(pixels);   // 源数据消费完毕（卷积/重采样都在构造期）
 
-        auto stagingAlloc = allocator.CreateBuffer(
-            total,
-            vk::BufferUsageFlagBits::eTransferSrc,
-            vk::MemoryPropertyFlagBits::eHostVisible |
-            vk::MemoryPropertyFlagBits::eHostCoherent);
-        std::memcpy(stagingAlloc.MappedPtr, faces.data(),
-                    static_cast<size_t>(total));
-        faces.clear();
-        faces.shrink_to_fit();   // CPU 侧临时数据尽早释放
+        // ---- 3. 上传双 cube + 共享采样器 ----
+        vk::Image envImage, irrImage;
+        auto envAlloc = UploadCube(device, allocator, envFaces, faceSize, envImage);
+        auto irrAlloc = UploadCube(device, allocator, irrFaces,
+                                   kIrradianceFaceSize, irrImage);
+        envFaces.clear(); envFaces.shrink_to_fit();
+        irrFaces.clear(); irrFaces.shrink_to_fit();
 
-        // ---- 4. cube image（eCubeCompatible，6 层，fp16）----
-        vk::ImageCreateInfo imageInfo;
-        imageInfo.setImageType(vk::ImageType::e2D)
-                 .setFormat(vk::Format::eR16G16B16A16Sfloat)
-                 .setExtent({faceSize, faceSize, 1})
-                 .setMipLevels(1)
-                 .setArrayLayers(kFaceCount)
-                 .setSamples(vk::SampleCountFlagBits::e1)
-                 .setTiling(vk::ImageTiling::eOptimal)
-                 .setFlags(vk::ImageCreateFlagBits::eCubeCompatible)
-                 .setUsage(vk::ImageUsageFlagBits::eTransferDst |
-                           vk::ImageUsageFlagBits::eSampled)
-                 .setSharingMode(vk::SharingMode::eExclusive)
-                 .setInitialLayout(vk::ImageLayout::eUndefined);
-
-        auto imageAlloc = allocator.CreateImage(imageInfo,
-            vk::MemoryPropertyFlagBits::eDeviceLocal);
-        const vk::Image image = imageAlloc.Image;
-
-        // ---- 5. cube view + 环境采样器（clamp-to-edge：cube 采样寻址由面自动处理）----
-        vk::ImageViewCreateInfo viewInfo;
-        viewInfo.setImage(image)
-                .setViewType(vk::ImageViewType::eCube)
-                .setFormat(vk::Format::eR16G16B16A16Sfloat)
-                .setSubresourceRange({
-                    vk::ImageAspectFlagBits::eColor, 0, 1, 0, kFaceCount});
-        vk::ImageView view = device.GetDevice().createImageView(viewInfo);
+        vk::ImageView envView = CreateCubeView(device, envImage);
+        vk::ImageView irrView = CreateCubeView(device, irrImage);
 
         vk::SamplerCreateInfo samplerInfo;
         samplerInfo.setMagFilter(vk::Filter::eLinear)
@@ -202,81 +349,52 @@ namespace Engine
                    .setMaxLod(0.0f);
         vk::Sampler sampler = device.GetDevice().createSampler(samplerInfo);
 
-        // ---- 6. 一次性命令缓冲：整 cube → TransferDst，逐面拷贝 → ShaderRead ----
-        vk::CommandPoolCreateInfo poolInfo;
-        poolInfo.setQueueFamilyIndex(device.GetGraphicsFamily())
-                .setFlags(vk::CommandPoolCreateFlagBits::eTransient);
-        vk::CommandPool pool = device.GetDevice().createCommandPool(poolInfo);
+        ENGINE_INFO("Environment cubemap: {} ({}x6 env + {}x6 irradiance, fp16)",
+                    path, faceSize, kIrradianceFaceSize);
 
-        vk::CommandBuffer cmd = device.GetDevice().allocateCommandBuffers(
-            vk::CommandBufferAllocateInfo{}
-                .setCommandPool(pool)
-                .setLevel(vk::CommandBufferLevel::ePrimary)
-                .setCommandBufferCount(1))[0];
-
-        cmd.begin(vk::CommandBufferBeginInfo{
-            vk::CommandBufferUsageFlagBits::eOneTimeSubmit});
-
-        // 6 层一起转 layout（subresource range layerCount = 6）
-        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(image)
-                  .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
-                                        0, 1, 0, kFaceCount});
-        cmd.pipelineBarrier2(
-            vk::DependencyInfo{}.setImageMemoryBarriers(toTransfer));
+        return Scope<Cubemap>(new Cubemap(device, allocator, envImage, envView,
+                                          irrImage, irrView, sampler,
+                                          envAlloc, irrAlloc, faceSize));
+    }
 
+    Scope<Cubemap> Cubemap::CreateNeutral(Device& device,
+                                          MemoryAllocator& allocator,
+                                          const glm::vec3& rgb)
+    {
+        // 1×1×6 常数 cube：env 与 irradiance 同值，描述符恒有效
+        const u16 px[4] = { FloatToHalf(rgb.r), FloatToHalf(rgb.g),
+                            FloatToHalf(rgb.b), 0x3C00 };
+        std::vector<u16> faces(kFaceCount * 4, 0);
         for (u32 f = 0; f < kFaceCount; ++f) {
-            vk::BufferImageCopy region;
-            region.setBufferOffset(faceBytes * f)
-                  .setBufferRowLength(0)
-                  .setBufferImageHeight(0)
-                  .setImageSubresource({vk::ImageAspectFlagBits::eColor,
-                                        0, f, 1})
-                  .setImageOffset({0, 0, 0})
-                  .setImageExtent({faceSize, faceSize, 1});
-            cmd.copyBufferToImage(stagingAlloc.Buffer, image,
-                                  vk::ImageLayout::eTransferDstOptimal, region);
+            std::memcpy(faces.data() + f * 4, px, sizeof(px));
         }
 
-        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(image)
-                .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
-                                      0, 1, 0, kFaceCount});
-        cmd.pipelineBarrier2(
-            vk::DependencyInfo{}.setImageMemoryBarriers(toShader));
+        vk::Image envImage, irrImage;
+        auto envAlloc = UploadCube(device, allocator, faces, 1, envImage);
+        auto irrAlloc = UploadCube(device, allocator, faces, 1, irrImage);
 
-        cmd.end();
+        vk::ImageView envView = CreateCubeView(device, envImage);
+        vk::ImageView irrView = CreateCubeView(device, irrImage);
 
-        // ---- 7. 提交并等待，清理 staging ----
-        vk::Fence fence = device.GetDevice().createFence({});
-        device.GetGraphicsQueue().submit(
-            vk::SubmitInfo{}.setCommandBuffers(cmd), fence);
-        if (device.GetDevice().waitForFences(fence, VK_TRUE, UINT64_MAX) !=
-            vk::Result::eSuccess) {
-            ENGINE_ERROR("Cubemap upload: waitForFences failed");
-        }
+        vk::SamplerCreateInfo samplerInfo;
+        samplerInfo.setMagFilter(vk::Filter::eNearest)
+                   .setMinFilter(vk::Filter::eNearest)
+                   .setMipmapMode(vk::SamplerMipmapMode::eNearest)
+                   .setAddressModeU(vk::SamplerAddressMode::eClampToEdge)
+                   .setAddressModeV(vk::SamplerAddressMode::eClampToEdge)
+                   .setAddressModeW(vk::SamplerAddressMode::eClampToEdge)
+                   .setAnisotropyEnable(VK_FALSE)
+                   .setBorderColor(vk::BorderColor::eIntOpaqueBlack)
+                   .setUnnormalizedCoordinates(VK_FALSE)
+                   .setCompareEnable(VK_FALSE)
+                   .setMinLod(0.0f)
+                   .setMaxLod(0.0f);
+        vk::Sampler sampler = device.GetDevice().createSampler(samplerInfo);
 
-        device.GetDevice().destroyFence(fence);
-        device.GetDevice().destroyCommandPool(pool);
-        allocator.Destroy(stagingAlloc);
-
-        ENGINE_INFO("Environment cubemap: {} ({}x6 {} faces, fp16)",
-                    path, faceSize, kFaceCount);
-
-        return Scope<Cubemap>(new Cubemap(device, allocator, image, view,
-                                          sampler, imageAlloc, faceSize));
+        ENGINE_INFO("Neutral cubemap created (constant fallback)");
+        return Scope<Cubemap>(new Cubemap(device, allocator, envImage, envView,
+                                          irrImage, irrView, sampler,
+                                          envAlloc, irrAlloc, 1));
     }
 
     Cubemap::~Cubemap()
@@ -284,10 +402,12 @@ namespace Engine
         const auto& vkDevice = m_Device.GetDevice();
 
         vkDevice.destroySampler(m_Sampler);
-        vkDevice.destroyImageView(m_View);
+        vkDevice.destroyImageView(m_EnvView);
+        vkDevice.destroyImageView(m_IrrView);
 
         // image 资源与显存经 VMA 成对释放
-        m_Allocator.Destroy(m_Allocation);
+        m_Allocator.Destroy(m_EnvAllocation);
+        m_Allocator.Destroy(m_IrrAllocation);
     }
 
 } // namespace Engine

Engine/src/Renderer/Passes/SceneGpu.cpp

@@ -4,6 +4,7 @@
 
 #include <Engine/Renderer/Passes/SceneGpu.h>
 #include <Engine/Renderer/Material.h>
+#include <Engine/Renderer/Cubemap.h>
 #include <Engine/Vulkan/Device.h>
 #include <Engine/Vulkan/DescriptorManager.h>
 #include <Engine/Vulkan/ShaderManager.h>
@@ -46,12 +47,16 @@ namespace Engine::Passes
         textureSetLayout = descriptors.CreateLayout(texBindings);
 
         // Set 2: CSM 级联阴影图（fragment 采样，binding = 级联序号）
+        // + 漫反射辐照度 cube（IBL 阶段②，binding 3）
         std::vector<DescriptorBinding> shadowBindings;
         for (u32 c = 0; c < kMaxCascades; ++c) {
             shadowBindings.push_back(
                 {c, vk::DescriptorType::eCombinedImageSampler,
                  vk::ShaderStageFlagBits::eFragment});
         }
+        shadowBindings.push_back(
+            {3, vk::DescriptorType::eCombinedImageSampler,
+             vk::ShaderStageFlagBits::eFragment});
         shadowSetLayout = descriptors.CreateLayout(shadowBindings);
 
         // Forward pipeline layout：三个 set；无 push（GPUDriven 经 SSBO）
@@ -157,4 +162,13 @@ namespace Engine::Passes
         }
     }
 
+    void SceneGpu::BindEnvironment(const Cubemap& env)
+    {
+        // irradiance view 全程稳定（不随 RG 重编译变化），绑一次即可
+        for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
+            m_Descriptors.UpdateTexture(shadowSets[slot], 3, shadowSampler,
+                                        env.GetIrradianceView());
+        }
+    }
+
 } // namespace Engine::Passes

Engine/src/Renderer/Renderer.cpp

@@ -74,12 +74,15 @@ namespace Engine
         m_FullscreenGpu = CreateScope<Passes::FullscreenGpu>(
             *m_Device, *m_DescriptorManager, *m_ShaderManager);
 
-        // 3.5 环境贴图（IBL 阶段①）：HDR 等距柱状 → cube。
-        //     文件缺失则跳过（SkyboxPass 自动退化为纯清屏，Forward 不受影响）
+        // 3.5 环境贴图（IBL）：HDR 等距柱状 → cube + irradiance 卷积。
+        //     文件缺失则创建中性常数兜底（描述符恒有效，渲染退化为常数环境）
         const std::string envPath = "assets/textures/venice_dawn_1_4k.hdr";
         if (std::filesystem::exists(envPath)) {
             m_Environment = Cubemap::CreateFromHdrEquirect(
                 *m_Device, *m_Allocator, envPath, 1024);
+        } else {
+            m_Environment = Cubemap::CreateNeutral(
+                *m_Device, *m_Allocator, glm::vec3(0.1f));
         }
 
         // 4. Swapchain（只管呈现；渲染目标统一归 RenderGraph）
@@ -107,6 +110,8 @@ namespace Engine
         for (auto& pass : m_Passes) {
             pass->OnTargetsRebuilt(*m_RenderGraph);
         }
+        // 环境辐照度绑定（view 全程稳定，绑一次）
+        m_SceneGpu->BindEnvironment(*m_Environment);
 
         // 9. GPU timestamp：每 pass 两枚（前后），按帧槽位分区
         m_QueriesPerFrame = static_cast<u32>(m_RenderGraph->GetPassCount()) * 2;