diff --git a/App/main.cpp b/App/main.cpp
index e6dbead..b221b4e 100644
--- a/App/main.cpp
+++ b/App/main.cpp
@@ -166,6 +166,13 @@ protected:
             renderer.SetBloomStrength(strength);
         }
 
+        // 调试视图：帧历史差异（静止=黑，运动边缘发光=链路正确）
+        bool temporalDiff = renderer.GetShowTemporalDiff();
+        if (ImGui::Checkbox("Temporal Diff View", &temporalDiff))
+        {
+            renderer.SetShowTemporalDiff(temporalDiff);
+        }
+
         // Pass 级 GPU 耗时（RG timestamp）
         ImGui::Separator();
         ImGui::Text("GPU Timings");
diff --git a/Engine/include/Engine/Core/Application.h b/Engine/include/Engine/Core/Application.h
index d644870..0a2462e 100644
--- a/Engine/include/Engine/Core/Application.h
+++ b/Engine/include/Engine/Core/Application.h
@@ -124,7 +124,7 @@ namespace Engine {
         u64 m_FrameCount = 0;
 
         // 帧率控制
-        bool m_VSync = false;        // 垂直同步（present 自带节奏）
+        bool m_VSync = true;        // 垂直同步（present 自带节奏）
         f32 m_TargetFps = 60.0f;    // 限帧目标（≥24）
 
         std::chrono::steady_clock::time_point m_LastFrameTime;
diff --git a/Engine/include/Engine/RenderGraph/RenderGraph.h b/Engine/include/Engine/RenderGraph/RenderGraph.h
index 870b169..0432e06 100644
--- a/Engine/include/Engine/RenderGraph/RenderGraph.h
+++ b/Engine/include/Engine/RenderGraph/RenderGraph.h
@@ -16,7 +16,6 @@
 //
 // 已知边界（后续演进点）：
 //   - 暂不支持异步 compute queue / 跨队列信号量
-//   - 暂不支持内存别名（需 VMA）与帧历史资源（TAA 等）
 //   - 同一 pass 内对同一资源"读+写"（input attachment / RWM）未建模
 //
 
@@ -98,6 +97,10 @@ namespace Engine::RG {
         vk::ImageView View(TextureHandle h) const;
         vk::Image Image(TextureHandle h) const;
 
+        // 帧历史资源：上一帧槽位的物理镜像（(slot + framesInFlight - 1) % F）。
+        // 采样描述符经 GetPhysicalView(h, 历史槽位) 预绑（同阴影图模式）。
+        u32 HistorySlot(u32 frameSlot) const;
+
         class RenderGraph* graph = nullptr;
     };
 
@@ -131,6 +134,10 @@ namespace Engine::RG {
         };
         std::vector<SlotState> slots;
 
+        // 是否存在帧历史读（Compile 期校验：历史资源必须有同帧正常使用，
+        // 否则没有物理镜像可读）
+        bool hasHistoryUse = false;
+
         // Compile 期状态链解算用（执行序上的"当前"状态）
         vk::ImageLayout chainLayout = vk::ImageLayout::eUndefined;
         vk::PipelineStageFlags2 chainStage{};
@@ -150,6 +157,7 @@ namespace Engine::RG {
             u32 resIndex;
             ResourceUse use;
             vk::ShaderStageFlags readStage;   // ShaderRead 时的消费 stage
+            bool history = false;             // 帧历史读：读上一帧槽位的物理镜像
         };
         std::vector<Use> uses;
         ExecuteFn exec;
@@ -161,10 +169,12 @@ namespace Engine::RG {
         //   steady = 该槽位已被使用过（从 resting 状态出发）
         //   first  = 该槽位首次使用（从 Undefined 出发）
         // imported 资源恒用 first（acquire 后 layout 语义上为 Undefined）
+        // history = 帧历史读（操作上一帧槽位的镜像，steady 变体带跨帧可见性同步）
         struct BarrierPair {
             u32 resIndex;
             vk::ImageMemoryBarrier2 steady;
             vk::ImageMemoryBarrier2 first;
+            bool history = false;
         };
         std::vector<BarrierPair> barriers;
     };
@@ -195,6 +205,11 @@ namespace Engine::RG {
         // 注意：物理镜像按槽位各一份，槽位变化/重编译后必须重新获取。
         vk::ImageView GetPhysicalView(TextureHandle h, u32 frameSlot) const;
 
+        // 帧历史槽位 = 上一帧使用的槽位：(slot - 1 + F) mod F
+        u32 HistorySlot(u32 frameSlot) const {
+            return (frameSlot + m_FramesInFlight - 1) % m_FramesInFlight;
+        }
+
         // ---- Pass 声明 ----
 
         class PassBuilder {
@@ -202,6 +217,12 @@ namespace Engine::RG {
             PassBuilder& Write(TextureHandle h, ResourceUse use);
             PassBuilder& Read(TextureHandle h, ResourceUse use,
                               vk::ShaderStageFlags stage = vk::ShaderStageFlagBits::eFragment);
+            // 帧历史读：读上一帧槽位的物理镜像（TAA/运动模糊/时域滤波用）。
+            //   - 不产生图内边：跨帧依赖由同队列提交序 + 历史 barrier 保证
+            //   - 资源必须有同帧正常使用（否则无物理镜像），Compile 期校验
+            //   - 采样描述符在 pass 内用 GetPhysicalView(h, HistorySlot(slot)) 预绑
+            PassBuilder& ReadHistory(TextureHandle h, ResourceUse use,
+                                     vk::ShaderStageFlags stage = vk::ShaderStageFlagBits::eFragment);
             PassBuilder& SideEffect();
 
         private:
diff --git a/Engine/include/Engine/Renderer/Passes/FullscreenGpu.h b/Engine/include/Engine/Renderer/Passes/FullscreenGpu.h
index 2c1c136..55073bb 100644
--- a/Engine/include/Engine/Renderer/Passes/FullscreenGpu.h
+++ b/Engine/include/Engine/Renderer/Passes/FullscreenGpu.h
@@ -37,6 +37,7 @@ namespace Engine {
             bool useACES = true;
             f32 bloomThreshold = 1.0f;   // HDR 域亮度阈值
             f32 bloomStrength = 0.8f;    // 叠加强度
+            bool showTemporalDiff = false;   // 调试视图：显示帧历史差异
         };
 
         class FullscreenGpu {
@@ -48,10 +49,11 @@ namespace Engine {
             FullscreenGpu(const FullscreenGpu&) = delete;
             FullscreenGpu& operator=(const FullscreenGpu&) = delete;
 
-            vk::DescriptorSetLayout setLayout;   // Set0 = {输入0, 输入1}
+            vk::DescriptorSetLayout setLayout;   // Set0 = {输入0, 输入1, 调试输入}
             vk::PipelineLayout layout;           // push 16B（FullscreenPush）
             vk::Sampler sampler;
             vk::ShaderEXT vert{}, extractFrag{}, blurFrag{}, tonemapFrag{};
+            vk::ShaderEXT diffFrag{};            // temporal_diff（帧历史调试）
 
             // 各 pass 自己持有 per-slot set；分配统一走这里
             vk::DescriptorSet AllocateSet();
diff --git a/Engine/include/Engine/Renderer/Passes/TemporalDiffPass.h b/Engine/include/Engine/Renderer/Passes/TemporalDiffPass.h
new file mode 100644
index 0000000..c20d565
--- /dev/null
+++ b/Engine/include/Engine/Renderer/Passes/TemporalDiffPass.h
@@ -0,0 +1,40 @@
+//
+// Created by atdunbg on 2026/9/23.
+//
+// TemporalDiffPass：帧历史链路调试视图。
+// 读 HDRColor 的帧历史镜像（ReadHistory）与当前帧，输出 |prev − cur| × 增益。
+// 静止画面纯黑、运动边缘发光 = 帧历史机制正确；噪声/残影错位 = 有 bug。
+// 也是 RenderGraph::ReadHistory 的参考实现（TAA 照此模式接入）。
+//
+
+#ifndef HELLOVULKAN_TEMPORALDIFFPASS_H
+#define HELLOVULKAN_TEMPORALDIFFPASS_H
+
+#include <Engine/Renderer/Passes/RenderPass.h>
+#include <Engine/Renderer/Passes/FullscreenGpu.h>
+#include <vulkan/vulkan.hpp>
+#include <vector>
+
+namespace Engine::Passes {
+
+    class TemporalDiffPass final : public RenderPass {
+    public:
+        TemporalDiffPass(FullscreenGpu& gpu, f32 gain,
+                         RG::TextureHandle hdrColor, RG::TextureHandle output);
+        ~TemporalDiffPass() override = default;
+
+        void Register(RG::RenderGraph& rg) override;
+        void OnTargetsRebuilt(RG::RenderGraph& rg) override;
+
+    private:
+        FullscreenGpu& m_Gpu;
+        f32 m_Gain;
+        RG::TextureHandle m_HDR;
+        RG::TextureHandle m_Output;
+
+        std::vector<vk::DescriptorSet> m_Sets;   // per-slot
+    };
+
+} // namespace Engine::Passes
+
+#endif //HELLOVULKAN_TEMPORALDIFFPASS_H
diff --git a/Engine/include/Engine/Renderer/Passes/TonemapPass.h b/Engine/include/Engine/Renderer/Passes/TonemapPass.h
index d6380c2..9a72d0d 100644
--- a/Engine/include/Engine/Renderer/Passes/TonemapPass.h
+++ b/Engine/include/Engine/Renderer/Passes/TonemapPass.h
@@ -20,7 +20,7 @@ namespace Engine::Passes {
     public:
         TonemapPass(FullscreenGpu& gpu, PostProcessParams& params,
                     RG::TextureHandle hdrColor, RG::TextureHandle bloomA,
-                    RG::TextureHandle backbuffer);
+                    RG::TextureHandle temporalDiff, RG::TextureHandle backbuffer);
         ~TonemapPass() override = default;
 
         void Register(RG::RenderGraph& rg) override;
@@ -31,6 +31,7 @@ namespace Engine::Passes {
         PostProcessParams& m_Params;
         RG::TextureHandle m_HDR;
         RG::TextureHandle m_A;
+        RG::TextureHandle m_TemporalDiff;   // 调试视图输入（当前帧槽位）
         RG::TextureHandle m_Backbuffer;
 
         std::vector<vk::DescriptorSet> m_Sets;   // per-slot
diff --git a/Engine/include/Engine/Renderer/Renderer.h b/Engine/include/Engine/Renderer/Renderer.h
index 08073d8..8a2932f 100644
--- a/Engine/include/Engine/Renderer/Renderer.h
+++ b/Engine/include/Engine/Renderer/Renderer.h
@@ -79,6 +79,9 @@ namespace Engine {
         void SetBloomThreshold(f32 v);
         f32 GetBloomStrength() const;
         void SetBloomStrength(f32 v);
+        // 调试视图：显示帧历史差异（TemporalDiff pass 产出）
+        bool GetShowTemporalDiff() const;
+        void SetShowTemporalDiff(bool b);
 
         // pass 级 GPU 耗时（最近完成帧，毫秒；与 GetPassGpuMs 下标对应）
         const std::vector<std::string>& GetPassNames() const { return m_PassNames; }
@@ -125,6 +128,7 @@ namespace Engine {
         RG::TextureHandle m_HDRColor{};   // Forward 输出（线性 HDR），Tonemap 归一
         RG::TextureHandle m_BloomA{};     // 半分辨率：提取 + 垂直模糊结果
         RG::TextureHandle m_BloomB{};     // 半分辨率：水平模糊中间结果
+        RG::TextureHandle m_TemporalDiff{};   // 半分辨率：帧历史差异（调试）
 
         // pass 共享的可调参数（Renderer 持有，App 层访问器转发）
         Scope<Passes::PostProcessParams> m_PostParams;
diff --git a/Engine/include/Engine/Vulkan/Swapchain.h b/Engine/include/Engine/Vulkan/Swapchain.h
index 1b94137..daeb6e6 100644
--- a/Engine/include/Engine/Vulkan/Swapchain.h
+++ b/Engine/include/Engine/Vulkan/Swapchain.h
@@ -61,7 +61,7 @@ namespace Engine
 
         vk::Format m_Format;
         vk::Extent2D m_Extent;
-        bool m_VSync = false;
+        bool m_VSync = true;
     };
 } // namespace Engine
 
diff --git a/Engine/src/RenderGraph/RenderGraph.cpp b/Engine/src/RenderGraph/RenderGraph.cpp
index bf0a9df..4203c4d 100644
--- a/Engine/src/RenderGraph/RenderGraph.cpp
+++ b/Engine/src/RenderGraph/RenderGraph.cpp
@@ -130,6 +130,23 @@ namespace Engine::RG {
         return *this;
     }
 
+    RenderGraph::PassBuilder& RenderGraph::PassBuilder::ReadHistory(
+        TextureHandle h, ResourceUse use, vk::ShaderStageFlags stage)
+    {
+        m_Graph.CheckMutable();
+        m_Graph.CheckHandle(h);
+        auto& entry = m_Graph.m_Textures[h.index];
+        if (entry.imported) {
+            // swapchain image 不按槽位持久存在，历史读无意义
+            throw std::runtime_error(
+                "RenderGraph: history read of imported resource '" + entry.name +
+                "' is not supported");
+        }
+        entry.hasHistoryUse = true;
+        m_Graph.m_Passes[m_PassIndex].uses.push_back({h.index, use, stage, true});
+        return *this;
+    }
+
     RenderGraph::PassBuilder& RenderGraph::PassBuilder::SideEffect()
     {
         m_Graph.CheckMutable();
@@ -265,6 +282,7 @@ namespace Engine::RG {
 
                 bool keep = pass.sideEffect;
                 for (const auto& use : pass.uses) {
+                    // 帧历史读照常参与：它消费了上一帧的产出，生产者必须存活
                     if (IsWriteUse(use.use) && readByKept[use.resIndex]) {
                         keep = true;
                     } else if (m_Textures[use.resIndex].imported) {
@@ -317,6 +335,11 @@ namespace Engine::RG {
                 auto& lw = lastWriter[use.resIndex];
                 auto& readers = pendingReaders[use.resIndex];
 
+                // 帧历史读不建边：它消费的是上一帧槽位的物理镜像（与本帧
+                // 任何写操作无资源竞争），跨帧依赖由同队列提交序 + 历史
+                // barrier（ALL_COMMANDS 可见性同步）保证
+                if (use.history) continue;
+
                 if (IsWriteUse(use.use)) {
                     for (u32 r : readers) addEdge(r, i);  // WAR
                     readers.clear();
@@ -363,6 +386,7 @@ namespace Engine::RG {
         for (u32 pos = 0; pos < m_ExecutionOrder.size(); ++pos) {
             const u32 passIdx = m_ExecutionOrder[pos];
             for (const auto& use : m_Passes[passIdx].uses) {
+                if (use.history) continue;   // 帧历史读不改变本帧状态链与生命周期
                 auto& entry = m_Textures[use.resIndex];
                 if (entry.firstPass < 0) entry.firstPass = static_cast<i32>(pos);
                 entry.lastPass = static_cast<i32>(pos);
@@ -387,6 +411,7 @@ namespace Engine::RG {
         // ---- Pass A：状态链推演 ----
         for (const u32 passIdx : m_ExecutionOrder) {
             for (const auto& use : m_Passes[passIdx].uses) {
+                if (use.history) continue;   // 历史读操作另一槽位的镜像，不入本帧链
                 auto& entry = m_Textures[use.resIndex];
                 entry.chainLayout = UseLayout(use.use);
                 entry.chainStage = UseDstStage(use);
@@ -416,6 +441,36 @@ namespace Engine::RG {
             for (const auto& use : pass.uses) {
                 auto& entry = m_Textures[use.resIndex];
 
+                // ---- 帧历史读：操作上一帧槽位的物理镜像，不进入本帧状态链。
+                //   first  = 该槽位镜像从未用过（图刚编译/重编译），old=Undefined
+                //   steady = 跨帧可见性同步：src=ALL_COMMANDS|MEMORY_WRITE
+                //            覆盖上一帧提交对该镜像的全部写入（同队列提交序
+                //            保证 barrier 执行时上一帧命令已全部完成，
+                //            ALL_COMMANDS 把可用性+可见性显式闭合），
+                //            old/new = resting layout（历史镜像静止在读布局）
+                if (use.history) {
+                    const vk::ImageLayout desiredLayout = UseLayout(use.use);
+                    const vk::PipelineStageFlags2 dstStage = UseDstStage(use);
+                    const vk::AccessFlags2 dstAccess = UseDstAccess(use);
+
+                    vk::ImageMemoryBarrier2 first;
+                    first.setSrcStageMask(vk::PipelineStageFlagBits2::eNone)
+                         .setSrcAccessMask(vk::AccessFlagBits2::eNone)
+                         .setDstStageMask(dstStage)
+                         .setDstAccessMask(dstAccess)
+                         .setOldLayout(vk::ImageLayout::eUndefined)
+                         .setNewLayout(desiredLayout)
+                         .setSubresourceRange({entry.aspect, 0, 1, 0, 1});
+
+                    vk::ImageMemoryBarrier2 steady = first;
+                    steady.setSrcStageMask(vk::PipelineStageFlagBits2::eAllCommands)
+                          .setSrcAccessMask(vk::AccessFlagBits2::eMemoryWrite)
+                          .setOldLayout(entry.restingLayout);
+
+                    pass.barriers.push_back({use.resIndex, steady, first, true});
+                    continue;   // 不推进 chain：本帧链只描述当前槽位镜像
+                }
+
                 const vk::ImageLayout desiredLayout = UseLayout(use.use);
                 const vk::PipelineStageFlags2 dstStage = UseDstStage(use);
                 const vk::AccessFlags2 dstAccess = UseDstAccess(use);
@@ -451,6 +506,17 @@ namespace Engine::RG {
                         steady = makeBarrier(entry.restingLayout,
                                              entry.restingStage,
                                              entry.restingAccess);
+                        if (entry.hasHistoryUse) {
+                            // 历史读打破了"同槽位两次进入之间有 fence 全等保"
+                            // 的前提（上一帧对该槽位镜像的历史读不在本槽位
+                            // fence 的保护范围内），src 升级为 ALL_COMMANDS
+                            // 覆盖跨帧的一切触碰
+                            steady.setSrcStageMask(
+                                       vk::PipelineStageFlagBits2::eAllCommands)
+                                  .setSrcAccessMask(
+                                       vk::AccessFlagBits2::eMemoryWrite |
+                                       vk::AccessFlagBits2::eMemoryRead);
+                        }
                     }
 
                     pass.barriers.push_back({use.resIndex, steady, first});
@@ -481,7 +547,16 @@ namespace Engine::RG {
     void RenderGraph::AllocatePhysical()
     {
         for (auto& entry : m_Textures) {
-            if (entry.imported || entry.firstPass < 0) continue;
+            if (entry.imported) continue;
+
+            // 帧历史校验：历史资源必须有同帧正常使用（否则既无产出也无镜像）
+            if (entry.hasHistoryUse && entry.firstPass < 0) {
+                throw std::runtime_error(
+                    "RenderGraph: history-read resource '" + entry.name +
+                    "' has no same-frame use (nothing to read history of)");
+            }
+
+            if (entry.firstPass < 0) continue;
 
             // 显式尺寸优先；否则 = 目标尺寸 × sizeScale（后处理降采样）
             vk::Extent2D extent{
@@ -649,21 +724,27 @@ namespace Engine::RG {
             for (const auto& bp : pass.barriers) {
                 auto& entry = m_Textures[bp.resIndex];
 
+                u32 imageSlot = frame.frameSlot;
                 bool useFirst;
                 if (entry.imported) {
                     useFirst = true; // 恒为 Undefined 起点
+                } else if (bp.history) {
+                    // 帧历史：操作上一帧槽位的物理镜像。
+                    // used 标记持久（跨帧），首帧走 first，之后走 steady
+                    imageSlot = HistorySlot(frame.frameSlot);
+                    useFirst = !entry.slots[imageSlot].used;
+                    entry.slots[imageSlot].used = true;
                 } else {
-                    auto& slot = entry.slots[frame.frameSlot];
-                    useFirst = !slot.used;
-                    slot.used = true;
+                    useFirst = !entry.slots[imageSlot].used;
+                    entry.slots[imageSlot].used = true;
                 }
 
                 vk::ImageMemoryBarrier2 b = useFirst ? bp.first : bp.steady;
                 // image 句柄在执行期绑定：imported 用当帧 acquire 的 image，
-                // created 用当前帧槽位的物理镜像（每个槽位各有一份）
+                // created 用当前帧槽位（历史 barrier 用上一帧槽位）的物理镜像
                 b.setImage(entry.imported
                                ? entry.importedImage
-                               : entry.physical[frame.frameSlot]->image);
+                               : entry.physical[imageSlot]->image);
                 m_ScratchBarriers.push_back(b);
             }
 
@@ -707,6 +788,11 @@ namespace Engine::RG {
         return entry.physical[frame.frameSlot]->image;
     }
 
+    u32 RenderGraphContext::HistorySlot(u32 frameSlot) const
+    {
+        return graph->HistorySlot(frameSlot);
+    }
+
     // ---------- 工具 / 统计 ----------
 
     TextureEntry& RenderGraph::GetEntry(TextureHandle h)
diff --git a/Engine/src/Renderer/Passes/FullscreenGpu.cpp b/Engine/src/Renderer/Passes/FullscreenGpu.cpp
index 4bae26f..bdf8f11 100644
--- a/Engine/src/Renderer/Passes/FullscreenGpu.cpp
+++ b/Engine/src/Renderer/Passes/FullscreenGpu.cpp
@@ -16,12 +16,16 @@ namespace Engine::Passes
                                  ShaderManager& shaders)
         : m_Device(device), m_Descriptors(descriptors)
     {
-        // Set0 = {HDR 颜色图, Bloom 图}，push = 16B 通用结构
+        // Set0 = {HDR 颜色图, Bloom 图, 调试输入}，push = 16B 通用结构
+        //（binding2 只被 tonemap 的调试视图使用；bloom 系 shader 静态
+        //  使用的 binding 是布局的子集，合法）
         std::vector<DescriptorBinding> bindings = {
             {0, vk::DescriptorType::eCombinedImageSampler,
              vk::ShaderStageFlagBits::eFragment},
             {1, vk::DescriptorType::eCombinedImageSampler,
              vk::ShaderStageFlagBits::eFragment},
+            {2, vk::DescriptorType::eCombinedImageSampler,
+             vk::ShaderStageFlagBits::eFragment},
         };
         setLayout = descriptors.CreateLayout(bindings);
 
@@ -59,13 +63,19 @@ namespace Engine::Passes
             vk::ShaderStageFlagBits::eFragment,
             std::vector<vk::DescriptorSetLayout>{setLayout},
             static_cast<u32>(sizeof(FullscreenPush)));
+        auto diffF = shaders.CreateFromFile(
+            shaderDir + "temporal_diff.frag.spv",
+            vk::ShaderStageFlagBits::eFragment,
+            std::vector<vk::DescriptorSetLayout>{setLayout},
+            static_cast<u32>(sizeof(FullscreenPush)));
 
         vert = vertShader.Handle;
         tonemapFrag = tonemapF.Handle;
         extractFrag = extractF.Handle;
         blurFrag = blurF.Handle;
+        diffFrag = diffF.Handle;
 
-        if (!vert || !tonemapFrag || !extractFrag || !blurFrag) {
+        if (!vert || !tonemapFrag || !extractFrag || !blurFrag || !diffFrag) {
             ENGINE_FATAL("FullscreenGpu: failed to create shader objects");
         }
     }
@@ -78,6 +88,7 @@ namespace Engine::Passes
         if (tonemapFrag) vkDevice.destroyShaderEXT(tonemapFrag);
         if (extractFrag) vkDevice.destroyShaderEXT(extractFrag);
         if (blurFrag) vkDevice.destroyShaderEXT(blurFrag);
+        if (diffFrag) vkDevice.destroyShaderEXT(diffFrag);
 
         vkDevice.destroySampler(sampler);
         vkDevice.destroyPipelineLayout(layout);
diff --git a/Engine/src/Renderer/Passes/TemporalDiffPass.cpp b/Engine/src/Renderer/Passes/TemporalDiffPass.cpp
new file mode 100644
index 0000000..d3750ac
--- /dev/null
+++ b/Engine/src/Renderer/Passes/TemporalDiffPass.cpp
@@ -0,0 +1,49 @@
+//
+// Created by atdunbg on 2026/9/23.
+//
+
+#include <Engine/Renderer/Passes/TemporalDiffPass.h>
+#include <Engine/Core/FrameContext.h>
+
+namespace Engine::Passes
+{
+    TemporalDiffPass::TemporalDiffPass(FullscreenGpu& gpu, f32 gain,
+                                       RG::TextureHandle hdrColor,
+                                       RG::TextureHandle output)
+        : m_Gpu(gpu), m_Gain(gain), m_HDR(hdrColor), m_Output(output)
+    {
+        m_Sets.resize(kMaxFramesInFlight);
+        for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
+            m_Sets[slot] = m_Gpu.AllocateSet();
+        }
+    }
+
+    void TemporalDiffPass::Register(RG::RenderGraph& rg)
+    {
+        // 帧历史读 + 当前帧读：同一虚拟资源的两个槽位镜像在同一个
+        // pass 里采样（TAA 的标准形态）。历史读不建图内边，跨帧同步
+        // 由 RG 的历史 barrier（ALL_COMMANDS 可见性）保证。
+        rg.AddPass("TemporalDiff",
+            [this](RG::RenderGraphContext& ctx) {
+                m_Gpu.RecordFullscreen(ctx, ctx.View(m_Output), m_Gpu.diffFrag,
+                                       m_Sets[ctx.frame.frameSlot],
+                                       FullscreenPush{0.0f, m_Gain, 0.0f, 0.0f});
+            })
+            .ReadHistory(m_HDR, RG::ResourceUse::ShaderRead,
+                         vk::ShaderStageFlagBits::eFragment)
+            .Read(m_HDR, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Write(m_Output, RG::ResourceUse::ColorAttachment);
+    }
+
+    void TemporalDiffPass::OnTargetsRebuilt(RG::RenderGraph& rg)
+    {
+        for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
+            // binding0 = 帧历史（上一帧槽位镜像），binding1 = 当前帧
+            m_Gpu.UpdateSetInput(m_Sets[slot], 0, rg, m_HDR,
+                                 rg.HistorySlot(slot));
+            m_Gpu.UpdateSetInput(m_Sets[slot], 1, rg, m_HDR, slot);
+        }
+    }
+
+} // namespace Engine::Passes
diff --git a/Engine/src/Renderer/Passes/TonemapPass.cpp b/Engine/src/Renderer/Passes/TonemapPass.cpp
index dcccf43..8356064 100644
--- a/Engine/src/Renderer/Passes/TonemapPass.cpp
+++ b/Engine/src/Renderer/Passes/TonemapPass.cpp
@@ -9,9 +9,11 @@ namespace Engine::Passes
 {
     TonemapPass::TonemapPass(FullscreenGpu& gpu, PostProcessParams& params,
                              RG::TextureHandle hdrColor, RG::TextureHandle bloomA,
+                             RG::TextureHandle temporalDiff,
                              RG::TextureHandle backbuffer)
         : m_Gpu(gpu), m_Params(params),
-          m_HDR(hdrColor), m_A(bloomA), m_Backbuffer(backbuffer)
+          m_HDR(hdrColor), m_A(bloomA), m_TemporalDiff(temporalDiff),
+          m_Backbuffer(backbuffer)
     {
         m_Sets.resize(kMaxFramesInFlight);
         for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
@@ -28,12 +30,17 @@ namespace Engine::Passes
                                        m_Sets[ctx.frame.frameSlot],
                                        FullscreenPush{m_Params.exposure,
                                                       m_Params.useACES ? 1.0f : 0.0f,
-                                                      m_Params.bloomStrength, 0.0f});
+                                                      m_Params.bloomStrength,
+                                                      m_Params.showTemporalDiff ? 1.0f : 0.0f});
             })
             .Write(m_Backbuffer, RG::ResourceUse::ColorAttachment)
             .Read(m_HDR, RG::ResourceUse::ShaderRead,
                   vk::ShaderStageFlagBits::eFragment)
             .Read(m_A, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            // shader 静态采样 binding2（调试视图开关在 push constant 里），
+            // 声明读否则本 pass 的产出无人消费会被剔除、镜像也不会分配
+            .Read(m_TemporalDiff, RG::ResourceUse::ShaderRead,
                   vk::ShaderStageFlagBits::eFragment);
     }
 
@@ -42,6 +49,7 @@ namespace Engine::Passes
         for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
             m_Gpu.UpdateSetInput(m_Sets[slot], 0, rg, m_HDR, slot);
             m_Gpu.UpdateSetInput(m_Sets[slot], 1, rg, m_A, slot);
+            m_Gpu.UpdateSetInput(m_Sets[slot], 2, rg, m_TemporalDiff, slot);
         }
     }
 
diff --git a/Engine/src/Renderer/Renderer.cpp b/Engine/src/Renderer/Renderer.cpp
index 4cb12d1..82211c3 100644
--- a/Engine/src/Renderer/Renderer.cpp
+++ b/Engine/src/Renderer/Renderer.cpp
@@ -16,6 +16,7 @@
 #include <Engine/Renderer/Passes/ForwardPass.h>
 #include <Engine/Renderer/Passes/BloomPass.h>
 #include <Engine/Renderer/Passes/TonemapPass.h>
+#include <Engine/Renderer/Passes/TemporalDiffPass.h>
 #include <Engine/Renderer/Passes/ImGuiPass.h>
 #include <Engine/Renderer/Passes/PresentPass.h>
 #include <Engine/Vulkan/Device.h>
@@ -143,6 +144,8 @@ namespace Engine
     void Renderer::SetBloomThreshold(f32 v) { m_PostParams->bloomThreshold = v; }
     f32 Renderer::GetBloomStrength() const { return m_PostParams->bloomStrength; }
     void Renderer::SetBloomStrength(f32 v) { m_PostParams->bloomStrength = v; }
+    bool Renderer::GetShowTemporalDiff() const { return m_PostParams->showTemporalDiff; }
+    void Renderer::SetShowTemporalDiff(bool b) { m_PostParams->showTemporalDiff = b; }
 
     u32 Renderer::GetMainInstanceCount() const
     {
@@ -194,6 +197,12 @@ namespace Engine
             "BloomB",
             RG::TextureDesc{vk::Format::eR16G16B16A16Sfloat, 0, 0, 0.5f});
 
+        // 帧历史差异（调试视图）：|上一帧 − 当前帧| × 增益。
+        // 帧历史机制（ReadHistory）的验证消费端，也是 TAA 的参考形态
+        m_TemporalDiff = m_RenderGraph->CreateTexture(
+            "TemporalDiff",
+            RG::TextureDesc{vk::Format::eR16G16B16A16Sfloat, 0, 0, 0.5f});
+
         // ---- pass 注册：执行序由拓扑排序决定，与注册顺序无关 ----
         // 要加新 pass：new 一个 RenderPass 子类，在这里 push 一行。
         FrameSceneData& frame = m_SceneRenderer->Frame();
@@ -201,10 +210,13 @@ namespace Engine
             *m_SceneGpu, frame, m_ShadowMap));
         m_Passes.push_back(CreateScope<Passes::ForwardPass>(
             *m_SceneGpu, frame, m_HDRColor, m_Depth, m_ShadowMap));
+        m_Passes.push_back(CreateScope<Passes::TemporalDiffPass>(
+            *m_FullscreenGpu, /*gain*/ 8.0f, m_HDRColor, m_TemporalDiff));
         m_Passes.push_back(CreateScope<Passes::BloomPass>(
             *m_FullscreenGpu, *m_PostParams, m_HDRColor, m_BloomA, m_BloomB));
         m_Passes.push_back(CreateScope<Passes::TonemapPass>(
-            *m_FullscreenGpu, *m_PostParams, m_HDRColor, m_BloomA, m_Backbuffer));
+            *m_FullscreenGpu, *m_PostParams, m_HDRColor, m_BloomA,
+            m_TemporalDiff, m_Backbuffer));
         m_Passes.push_back(CreateScope<Passes::ImGuiPass>(
             &m_ImGuiLayer, m_Backbuffer));
         m_Passes.push_back(CreateScope<Passes::PresentPass>(m_Backbuffer));
diff --git a/shaders/temporal_diff.frag b/shaders/temporal_diff.frag
new file mode 100644
index 0000000..2433ccd
--- /dev/null
+++ b/shaders/temporal_diff.frag
@@ -0,0 +1,28 @@
+#version 450
+
+// TemporalDiff：帧历史链路调试视图。
+// 输出 |上一帧 HDR − 当前帧 HDR| × 增益：
+//   - 静止画面应接近纯黑（两帧内容一致）
+//   - 运动的实体边缘发光（时域差异可见）
+//   - 全屏噪声 / 错位残影 = 帧历史 barrier 或槽位镜像绑定错误
+// 这是 RenderGraph 帧历史资源（ReadHistory）的验证消费端。
+
+layout(location = 0) in vec2 fragUV;
+layout(location = 0) out vec4 outColor;
+
+layout(set = 0, binding = 0) uniform sampler2D prevImage;   // 帧历史（上一帧槽位镜像）
+layout(set = 0, binding = 1) uniform sampler2D currImage;   // 当前帧
+
+// 全屏 pass 通用 push（16B，程序级统一声明，VUID-08878）
+layout(push_constant) uniform PushConstants {
+    float exposure;   // 与其它全屏 pass 布局一致（本 pass 不用）
+    float gain;       // 差异增益
+    float param1;
+    float param2;
+} pc;
+
+void main() {
+    vec3 prev = texture(prevImage, fragUV).rgb;
+    vec3 curr = texture(currImage, fragUV).rgb;
+    outColor = vec4(abs(prev - curr) * pc.gain, 1.0);
+}
diff --git a/shaders/tonemap.frag b/shaders/tonemap.frag
index 8e98a3c..0b95f22 100644
--- a/shaders/tonemap.frag
+++ b/shaders/tonemap.frag
@@ -10,13 +10,14 @@ layout(location = 0) out vec4 outColor;
 
 layout(set = 0, binding = 0) uniform sampler2D hdrImage;
 layout(set = 0, binding = 1) uniform sampler2D bloomImage;
+layout(set = 0, binding = 2) uniform sampler2D temporalDiffImage;   // 帧历史差异（调试视图）
 
 // 全屏 pass 通用 push（16B，程序级统一声明，VUID-08878）
 layout(push_constant) uniform PushConstants {
     float exposure;      // 曝光系数
     float aces;          // >0.5 启用 ACES 曲线，否则纯 gamma
     float bloomStrength; // bloom 叠加强度（0 = 关闭）
-    float _pad0;
+    float debugView;     // >0.5 显示帧历史差异视图（TemporalDiff）
 } pc;
 
 vec3 acesFilm(vec3 x) {
@@ -24,6 +25,13 @@ vec3 acesFilm(vec3 x) {
 }
 
 void main() {
+    // 调试视图：帧历史差异（静止=黑，运动边缘发光）
+    if (pc.debugView > 0.5) {
+        vec3 d = clamp(texture(temporalDiffImage, fragUV).rgb, 0.0, 1.0);
+        outColor = vec4(pow(d, vec3(1.0 / 2.2)), 1.0);
+        return;
+    }
+
     // bloom 在 HDR 域叠加（半分辨率模糊图线性采样上采样）
     vec4 hdrTex = texture(hdrImage, fragUV);
     vec3 hdr = hdrTex.rgb + texture(bloomImage, fragUV).rgb * pc.bloomStrength;