diff --git a/Engine/include/Engine/Renderer/Mesh.h b/Engine/include/Engine/Renderer/Mesh.h
index f26e068..f04bbc1 100644
--- a/Engine/include/Engine/Renderer/Mesh.h
+++ b/Engine/include/Engine/Renderer/Mesh.h
@@ -47,6 +47,8 @@ namespace Engine
         {
             return m_Lods[lod < kMeshLodCount ? lod : kMeshLodCount - 1].indexCount;
         }
+        // 全局连续槽位（SceneRenderer 数组化 batch 查找用，避免逐实例 hash）
+        u32 GetSlot() const { return m_Slot; }
 
         // 局部空间包围球（构造时从顶点算出），视锥剔除用
         struct Bounds {
@@ -66,6 +68,7 @@ namespace Engine
         Scope<Buffer> m_IndexBuffer;
         std::array<LodRange, kMeshLodCount> m_Lods{};
         Material* m_Material;
+        u32 m_Slot = 0;
         Bounds m_Bounds;
 
         u64 m_VertexBytes = 0;
diff --git a/Engine/include/Engine/Renderer/Model.h b/Engine/include/Engine/Renderer/Model.h
index ad55e91..a7ef320 100644
--- a/Engine/include/Engine/Renderer/Model.h
+++ b/Engine/include/Engine/Renderer/Model.h
@@ -81,6 +81,9 @@ namespace Engine
         Model& operator=(const Model&) = delete;
 
         const std::vector<Scope<Mesh>>& GetMeshes() const { return m_Meshes; }
+        // 全模型联合包围球（各 mesh bounds 的并集，局部空间）：
+        // 实体级粗剔除用——整个模型在视锥外直接跳过全部 mesh 循环
+        const Mesh::Bounds& GetBounds() const { return m_Bounds; }
 
     private:
         Model(Device& device, MemoryAllocator& allocator,
@@ -103,6 +106,9 @@ namespace Engine
                                       bool isGltf);
         void InitDefaultTextures();
 
+        // 聚合各 mesh bounds 为全模型联合包围球（实体级粗剔除）
+        void RebuildUnionBounds();
+
         Texture* GetOrLoadTexture(const std::string& path, bool srgb);
 
         // 材质缓存入口：同贴图组合+因子共享同一 Material（一个描述符集），
@@ -124,6 +130,7 @@ namespace Engine
         std::array<Scope<Texture>, Material::kSlotCount> m_DefaultTextures;
 
         std::vector<Scope<Mesh>> m_Meshes;
+        Mesh::Bounds m_Bounds;   // mesh 建完后聚合（实体级粗剔除）
         std::string m_ModelPath;
         bool m_IsGltf = false;   // glTF 家族：MR 合图语义 + 不翻转 UV
     };
diff --git a/Engine/include/Engine/Renderer/Renderer.h b/Engine/include/Engine/Renderer/Renderer.h
index c51f18d..ec9ce7d 100644
--- a/Engine/include/Engine/Renderer/Renderer.h
+++ b/Engine/include/Engine/Renderer/Renderer.h
@@ -111,6 +111,7 @@ namespace Engine {
         u32 GetMainInstanceCount() const;
         u32 GetShadowInstanceCount() const;
         bool GetInstancesTruncated() const;   // 实例预算溢出（遥测红字）
+        f32 GetCollectMs() const;             // CPU 收集+剔除段耗时
 
         // 垂直同步开关（present mode 切换会重建 swapchain），
         // 只应在帧 UI 构建段调用（Render 之前）
diff --git a/Engine/include/Engine/Renderer/SceneRenderer.h b/Engine/include/Engine/Renderer/SceneRenderer.h
index 5c0aaec..8b38d75 100644
--- a/Engine/include/Engine/Renderer/SceneRenderer.h
+++ b/Engine/include/Engine/Renderer/SceneRenderer.h
@@ -21,8 +21,10 @@
 #define HELLOVULKAN_SCENERENDERER_H
 
 #include <Engine/Core/Types.h>
+#include <Engine/Renderer/Mesh.h>
 #include <Engine/Renderer/SceneTypes.h>
 #include <vulkan/vulkan.hpp>
+#include <array>
 #include <unordered_map>
 
 namespace Engine {
@@ -67,23 +69,26 @@ namespace Engine {
 
         // ---- 剔除中间量（每帧 clear 复用，拼接进 m_Frame 后保留容量）----
         struct ViewScratch {
+            // cur + prev 合并写：两次随机堆访问 → 一次（cache miss 减半）
+            struct InstanceMatrices {
+                glm::mat4 cur;
+                glm::mat4 prev;
+            };
             CullResult result;
             // 每 batch 自己的实例列表（按 mesh 分组）。GPU 端 batch 语义 =
             // firstInstance + 连续区间，实体序收集不满足（多实体共享 mesh
             // 时实例分散各处），扁平化成连续数组推迟到 appendView
-            std::vector<std::vector<glm::mat4>> batchWorlds;
-            std::vector<std::vector<glm::mat4>> batchPrevWorlds;
-            // key = (Mesh*, LOD)：同 mesh 不同 LOD 索引区间不同，
-            // 必须分开成 batch（低 8 位 LOD，高位指针，指针 8 字节对齐安全）
-            std::unordered_map<u64, u32> lookup;
+            std::vector<std::vector<InstanceMatrices>> batchMats;
+            // batch 查找：数组寻址（下标 = Mesh::GetSlot()，值 = 每 LOD 的
+            // batchIdx+1，0 = 未用）——逐实例 hash 在大场景是毫秒级开销
+            std::vector<std::array<u32, kMeshLodCount>> lookup;
 
             void Clear()
             {
                 result.batches.clear();
                 result.instanceCount = 0;
-                batchWorlds.clear();
-                batchPrevWorlds.clear();
-                lookup.clear();
+                batchMats.clear();
+                for (auto& e : lookup) e.fill(0);   // 容量保留，置零复用
             }
         };
         ViewScratch m_MainScratch;
diff --git a/Engine/src/Renderer/Mesh.cpp b/Engine/src/Renderer/Mesh.cpp
index 8b0b07a..08f4bae 100644
--- a/Engine/src/Renderer/Mesh.cpp
+++ b/Engine/src/Renderer/Mesh.cpp
@@ -10,16 +10,24 @@
 #include <Engine/Core/Log.h>
 
 #include <limits>
+#include <atomic>
 
 #include "Engine/Core/Stats.h"
 
 namespace Engine
 {
+    namespace {
+        // 全局连续 mesh 槽位：SceneRenderer 的 batch 查找用数组寻址
+        //（每实例每视图一次 hash 的代价在大场景下是 5ms 级）
+        std::atomic<u32> s_nextMeshSlot{0};
+    }
+
     Mesh::Mesh(Device& device, MemoryAllocator& allocator,
                const std::vector<Vertex>& vertices,
                const std::vector<std::vector<u32>>& lods,
                Material* material)
         : m_Material(material)
+        , m_Slot(s_nextMeshSlot.fetch_add(1, std::memory_order_relaxed))
     {
         if (vertices.empty() || lods.empty() || lods[0].empty())
         {
diff --git a/Engine/src/Renderer/Model.cpp b/Engine/src/Renderer/Model.cpp
index 9be1986..ebf174e 100644
--- a/Engine/src/Renderer/Model.cpp
+++ b/Engine/src/Renderer/Model.cpp
@@ -45,6 +45,23 @@ Model::Model(Device& device, MemoryAllocator& allocator,
 
 Model::~Model() = default;
 
+void Model::RebuildUnionBounds()
+{
+    if (m_Meshes.empty()) {
+        m_Bounds = {};
+        return;
+    }
+    // 各 mesh 包围球 → AABB 并集 → 外接球（保守，宁多画不漏画）
+    glm::vec3 mn(1e9f), mx(-1e9f);
+    for (const auto& mesh : m_Meshes) {
+        const auto& b = mesh->GetBounds();
+        mn = glm::min(mn, b.center - glm::vec3(b.radius));
+        mx = glm::max(mx, b.center + glm::vec3(b.radius));
+    }
+    m_Bounds.center = (mn + mx) * 0.5f;
+    m_Bounds.radius = 0.5f * glm::length(mx - mn);
+}
+
 void Model::InitDefaultTextures()
 {
     // 各材质槽位的兜底纹理（1x1，开销可忽略）：
@@ -91,6 +108,7 @@ Model::Model(Device& device, MemoryAllocator& allocator,
 
     m_Meshes.push_back(CreateScope<Mesh>(
         m_Device, m_Allocator, vertices, lodIndices, mat));
+    RebuildUnionBounds();
 }
 
 Scope<Model> Model::CreatePlane(Device& device, MemoryAllocator& allocator,
@@ -165,6 +183,7 @@ Scope<Model> Model::FromParsed(Device& device, MemoryAllocator& allocator,
     }
     ENGINE_INFO("Model loaded: {} ({} meshes, {} textures)",
                 parsed.path, model->m_Meshes.size(), model->m_TextureCache.size());
+    model->RebuildUnionBounds();
     return model;
 }
 
@@ -184,6 +203,7 @@ void Model::LoadModel(const std::string& path)
     }
     ENGINE_INFO("Model loaded: {} ({} meshes, {} textures)",
                 path, m_Meshes.size(), m_TextureCache.size());
+    RebuildUnionBounds();
 }
 
 void Model::ProcessNode(const aiNode* node, const aiScene* scene,
diff --git a/Engine/src/Renderer/Renderer.cpp b/Engine/src/Renderer/Renderer.cpp
index aa9c67c..d358862 100644
--- a/Engine/src/Renderer/Renderer.cpp
+++ b/Engine/src/Renderer/Renderer.cpp
@@ -225,6 +225,11 @@ namespace Engine
         return m_SceneRenderer->Frame().instancesTruncated;
     }
 
+    f32 Renderer::GetCollectMs() const
+    {
+        return m_SceneRenderer->GetCollectMs();
+    }
+
     u32 Renderer::GetShadowInstanceCount() const
     {
         return m_SceneRenderer->GetShadowInstanceCount();
diff --git a/Engine/src/Renderer/SceneRenderer.cpp b/Engine/src/Renderer/SceneRenderer.cpp
index 31d23a0..00c54df 100644
--- a/Engine/src/Renderer/SceneRenderer.cpp
+++ b/Engine/src/Renderer/SceneRenderer.cpp
@@ -291,6 +291,23 @@ namespace Engine
                     glm::max(glm::length(glm::vec3(t.World()[1])),
                              glm::length(glm::vec3(t.World()[2]))));
 
+                // ---- 实体级粗剔除（联合包围球）：四视图全部看不到 →
+                // 跳过整个 mesh 循环（多 mesh 模型的大头优化）----
+                const Mesh::Bounds& mb = model->GetBounds();
+                const glm::vec3 modelCenter =
+                    glm::vec3(t.World() * glm::vec4(mb.center, 1.0f));
+                const f32 modelRadius = mb.radius * worldScale;
+                const bool mainVis =
+                    mainFrustum.IntersectsSphere(modelCenter, modelRadius);
+                bool cascVis[kMaxCascades];
+                bool anyVis = mainVis;
+                for (u32 c = 0; c < kMaxCascades; ++c) {
+                    cascVis[c] =
+                        cascFrustum[c].IntersectsSphere(modelCenter, modelRadius);
+                    anyVis |= cascVis[c];
+                }
+                if (!anyVis) return;
+
                 for (auto& mesh : model->GetMeshes()) {
                     const auto& bounds = mesh->GetBounds();
                     const glm::vec3 worldCenter =
@@ -317,31 +334,32 @@ namespace Engine
                             instanceTruncated = true;
                             return;
                         }
-                        // lookup 值 = batches 索引（此前误存 worlds 索引：
-                        // 单实体时两序号恰好恒等，多实体共享 mesh 后分叉
-                        // → batches[worldsIdx] 越界）。batch.first（GPU
-                        // firstInstance 基址）才是 worlds 索引，语义别混
-                        const u64 key =
-                            (reinterpret_cast<u64>(mesh.get()) << 8) | lod;
-                        const auto [it, inserted] = s.lookup.emplace(
-                            key, static_cast<u32>(s.result.batches.size()));
-                        if (inserted) {
+                        // batch 查找 = 数组寻址（Mesh::GetSlot 槽位 + LOD 槽）
+                        // batchIdx 存储 +1 偏移（0 = 该槽位尚无 batch）
+                        const u32 slot = mesh->GetSlot();
+                        if (slot >= s.lookup.size()) {
+                            s.lookup.resize(slot + 1);
+                        }
+                        auto& lodSlots = s.lookup[slot];
+                        u32 batchIdx = lodSlots[lod];
+                        if (batchIdx == 0) {
+                            batchIdx = static_cast<u32>(s.result.batches.size()) + 1;
+                            lodSlots[lod] = batchIdx;
                             s.result.batches.push_back({mesh.get(), 0, 0, lod});
-                            s.batchWorlds.emplace_back();
-                            s.batchPrevWorlds.emplace_back();
+                            s.batchMats.emplace_back();
                         }
                         // 实例归入自己的 batch（mesh 分组），保证同 batch
                         // 实例在扁平化后连续——firstInstance 语义的前提
-                        const u32 batchIdx = it->second;
-                        s.result.batches[batchIdx].count++;
-                        s.batchWorlds[batchIdx].push_back(t.World());
-                        s.batchPrevWorlds[batchIdx].push_back(prevWorld);
+                        const u32 bi = batchIdx - 1;
+                        s.result.batches[bi].count++;
+                        s.batchMats[bi].push_back({t.World(), prevWorld});
                         s.result.instanceCount++;
                         ++totalInstances;
                     };
 
                     tryAppend(m_MainScratch, mainFrustum, mainLod);
                     for (u32 c = 0; c < kMaxCascades; ++c) {
+                        if (!cascVis[c]) continue;   // 粗剔除已判不可见
                         // 级联降级：越远的级联覆盖范围越大，用越低的 LOD
                         //（阴影本来就是低频信息，误差可接受）
                         tryAppend(m_CascScratch[c], cascFrustum[c],
@@ -360,13 +378,10 @@ namespace Engine
                 b.first = base;          // 扁平化后的连续区间基址
                 base += b.count;
                 dst.batches.push_back(b);
-                m_Frame.instanceModels.insert(m_Frame.instanceModels.end(),
-                                              s.batchWorlds[i].begin(),
-                                              s.batchWorlds[i].end());
-                m_Frame.prevInstanceModels.insert(
-                    m_Frame.prevInstanceModels.end(),
-                    s.batchPrevWorlds[i].begin(),
-                    s.batchPrevWorlds[i].end());
+                for (const auto& m : s.batchMats[i]) {
+                    m_Frame.instanceModels.push_back(m.cur);
+                    m_Frame.prevInstanceModels.push_back(m.prev);
+                }
             }
             dst.instanceCount = s.result.instanceCount;
         };