diff --git a/src/data/store/ChunkedVolumeStore.cpp b/src/data/store/ChunkedVolumeStore.cpp
index 12a55f2..76a651f 100644
--- a/src/data/store/ChunkedVolumeStore.cpp
+++ b/src/data/store/ChunkedVolumeStore.cpp
@@ -88,6 +88,38 @@ json brickIndexJson(std::int64_t offset, std::int64_t clen, int bw, int bh,
               {"bd", bd}};
 }
 
+// 2×2×2 平均降采样核（buildPyramid 与 buildPyramidStreaming 共用，DRY）：
+// 给定源维度 (snx,sny,snz) 与「读源体素 (si,sj,sk)」回调，算出 dst 体素 (i,j,k)
+// 的降采样值——2×2×2 邻域非 blank 平均（round），全 blank → kBlank。越界邻域跳过。
+template <typename SrcAt>
+std::int16_t downsampleVoxel(int i, int j, int k, int snx, int sny, int snz,
+                             SrcAt&& srcAt) {
+  long sum = 0;
+  int cnt = 0;
+  for (int dk = 0; dk < 2; ++dk) {
+    const int sk = 2 * k + dk;
+    if (sk >= snz) continue;
+    for (int dj = 0; dj < 2; ++dj) {
+      const int sj = 2 * j + dj;
+      if (sj >= sny) continue;
+      for (int di = 0; di < 2; ++di) {
+        const int si = 2 * i + di;
+        if (si >= snx) continue;
+        const std::int16_t v = srcAt(si, sj, sk);
+        if (v == kBlank) continue;
+        sum += v;
+        ++cnt;
+      }
+    }
+  }
+  if (cnt == 0) return kBlank;
+  const long avg = std::lround(static_cast<double>(sum) / cnt);
+  return static_cast<std::int16_t>(avg);
+}
+
+// 降采样后某轴维度：ceil(n/2)，至少 1（buildPyramid 与流式共用）。
+int halfDim(int n) { return std::max(1, (n + 1) / 2); }
+
 // 从体中拷出一块的 int16（块内 i 最快、k 最慢，与体一致）。
 std::vector<std::int16_t> sliceBrick(const geopro::core::ScalarVolumeI16& vol,
                                      int bx, int by, int bz, int brick,
@@ -416,41 +448,19 @@ void ChunkedVolumeStore::buildPyramid(int levels) {
   for (int L = 1; L < target; ++L) {
     const DenseLevel& src = dense.back();
     DenseLevel dst;
-    dst.nx = std::max(1, (src.nx + 1) / 2);
-    dst.ny = std::max(1, (src.ny + 1) / 2);
-    dst.nz = std::max(1, (src.nz + 1) / 2);
+    dst.nx = halfDim(src.nx);
+    dst.ny = halfDim(src.ny);
+    dst.nz = halfDim(src.nz);
     dst.vox.assign(static_cast<std::size_t>(dst.nx) * dst.ny * dst.nz, kBlank);
-    for (int k = 0; k < dst.nz; ++k) {
-      for (int j = 0; j < dst.ny; ++j) {
-        for (int i = 0; i < dst.nx; ++i) {
-          // 2×2×2 邻域非 blank 平均（round），全 blank → kBlank。
-          long sum = 0;
-          int cnt = 0;
-          for (int dk = 0; dk < 2; ++dk) {
-            const int sk = 2 * k + dk;
-            if (sk >= src.nz) continue;
-            for (int dj = 0; dj < 2; ++dj) {
-              const int sj = 2 * j + dj;
-              if (sj >= src.ny) continue;
-              for (int di = 0; di < 2; ++di) {
-                const int si = 2 * i + di;
-                if (si >= src.nx) continue;
-                const std::int16_t v =
-                    src.vox[idxAt(src.nx, src.ny, si, sj, sk)];
-                if (v == kBlank) continue;
-                sum += v;
-                ++cnt;
-              }
-            }
-          }
-          if (cnt > 0) {
-            const long avg = std::lround(static_cast<double>(sum) / cnt);
-            dst.vox[idxAt(dst.nx, dst.ny, i, j, k)] =
-                static_cast<std::int16_t>(avg);
-          }
-        }
-      }
-    }
+    const auto srcAt = [&](int si, int sj, int sk) -> std::int16_t {
+      return src.vox[idxAt(src.nx, src.ny, si, sj, sk)];
+    };
+    for (int k = 0; k < dst.nz; ++k)
+      for (int j = 0; j < dst.ny; ++j)
+        for (int i = 0; i < dst.nx; ++i)
+          // 2×2×2 邻域非 blank 平均（round），全 blank → kBlank（共用核 DRY）。
+          dst.vox[idxAt(dst.nx, dst.ny, i, j, k)] =
+              downsampleVoxel(i, j, k, src.nx, src.ny, src.nz, srcAt);
     dense.push_back(std::move(dst));
   }
 
@@ -513,12 +523,16 @@ void ChunkedVolumeStore::buildPyramid(int levels) {
     rebuilt.push_back(std::move(lv));
   }
 
+  finalizePyramidMeta(std::move(rebuilt));
+}
+
+void ChunkedVolumeStore::finalizePyramidMeta(std::vector<Level> rebuilt) {
   levels_ = std::move(rebuilt);
   levelCount_ = static_cast<int>(levels_.size());
   // 兼容字段：level 0 索引（含新算的 min/max）回写 bricks_。
   bricks_ = levels_[0].bricks;
 
-  // 3) 重写 meta.json：保留所有原字段（含原 bricks），追加/覆盖 levels 数组。
+  // 重写 meta.json：保留所有原字段（含原 bricks），追加/覆盖 levels 数组。
   const fs::path metaPath = fs::path(dir_) / kMetaFile;
   json meta;
   {
@@ -565,6 +579,134 @@ void ChunkedVolumeStore::buildPyramid(int levels) {
   if (!metaOut) throw std::runtime_error("ChunkedVolumeStore: meta.json write failed");
 }
 
+void ChunkedVolumeStore::buildPyramidStreaming(int levels) {
+  // 参数语义同 buildPyramid：levels = 最高级索引，总级数 = levels+1，levels<=0 仅
+  // level 0。区别：各降采样级**逐块**由盘上 level L-1 的 ≤2×2×2 邻块降采样得到，
+  // 不重组整卷——任意时刻只持几个 L-1 邻块 + 一个 L 块。
+  const int target = levels < 0 ? 1 : levels + 1;
+  const int brick = meta_.brick;
+
+  std::vector<Level> rebuilt;
+  rebuilt.reserve(static_cast<std::size_t>(target));
+
+  // level 0：数据文件（data.bin）不变，复用现有 offset/clen，仅补每块 min/max。
+  // 内存有界：逐块读盘算 range，不重组整卷。
+  {
+    Level lv0 = levels_[0];  // 拷 nx/ny/nz/块数/dataFile/索引
+    for (int bz = 0; bz < lv0.bz; ++bz)
+      for (int by = 0; by < lv0.by; ++by)
+        for (int bx = 0; bx < lv0.bx; ++bx) {
+          BrickEntry& be =
+              lv0.bricks[static_cast<std::size_t>(brickIndexAt(lv0, bx, by, bz))];
+          const auto rng = computeRange(readBrickFrom(levels_[0], bx, by, bz));
+          be.vmin = rng.first;
+          be.vmax = rng.second;
+          be.hasRange = true;
+        }
+    rebuilt.push_back(std::move(lv0));
+  }
+
+  // 降采样级 L=1..target-1：源 = rebuilt.back()（盘上 level L-1）。
+  for (int L = 1; L < target; ++L) {
+    const Level& src = rebuilt.back();  // 提供 dims/块数/dataFile，readBrickFrom 从盘读
+
+    Level lv;
+    lv.nx = halfDim(src.nx);
+    lv.ny = halfDim(src.ny);
+    lv.nz = halfDim(src.nz);
+    lv.bx = ceilDiv(lv.nx, brick);
+    lv.by = ceilDiv(lv.ny, brick);
+    lv.bz = ceilDiv(lv.nz, brick);
+    lv.dataFile = levelDataFile(L);
+
+    std::ofstream data((fs::path(dir_) / lv.dataFile).string(),
+                       std::ios::binary | std::ios::trunc);
+    if (!data)
+      throw std::runtime_error("ChunkedVolumeStore: cannot open level data for write");
+    std::int64_t offset = 0;
+
+    // 逐 L 块（固定顺序 bz 最慢、bx 最快，与 write/buildPyramid 一致）。
+    for (int bz = 0; bz < lv.bz; ++bz) {
+      for (int by = 0; by < lv.by; ++by) {
+        for (int bx = 0; bx < lv.bx; ++bx) {
+          const int bw = std::min(brick, lv.nx - bx * brick);
+          const int bh = std::min(brick, lv.ny - by * brick);
+          const int bd = std::min(brick, lv.nz - bz * brick);
+          // 该 L 块 dst 体素全局坐标基点。
+          const int i0 = bx * brick, j0 = by * brick, k0 = bz * brick;
+
+          // 覆盖的 L-1 体素子区域 = [2*i0, 2*(i0+bw)) 截到 src 维度。
+          const int sLoI = 2 * i0, sLoJ = 2 * j0, sLoK = 2 * k0;
+          const int sHiI = std::min(src.nx, 2 * (i0 + bw));
+          const int sHiJ = std::min(src.ny, 2 * (j0 + bh));
+          const int sHiK = std::min(src.nz, 2 * (k0 + bd));
+
+          // 落在的 L-1 邻块范围（≤2×2×2）。读入小缓存（按相对块索引存）。
+          const int sbi0 = sLoI / brick, sbi1 = (sHiI - 1) / brick;
+          const int sbj0 = sLoJ / brick, sbj1 = (sHiJ - 1) / brick;
+          const int sbk0 = sLoK / brick, sbk1 = (sHiK - 1) / brick;
+          const int nbi = sbi1 - sbi0 + 1;
+          const int nbj = sbj1 - sbj0 + 1;
+          const int nbk = sbk1 - sbk0 + 1;
+          std::vector<std::vector<std::int16_t>> cache(
+              static_cast<std::size_t>(nbi) * nbj * nbk);
+          for (int kk = 0; kk < nbk; ++kk)
+            for (int jj = 0; jj < nbj; ++jj)
+              for (int ii = 0; ii < nbi; ++ii)
+                cache[(static_cast<std::size_t>(kk) * nbj + jj) * nbi + ii] =
+                    readBrickFrom(src, sbi0 + ii, sbj0 + jj, sbk0 + kk);
+
+          // 从邻块缓存读源体素 (si,sj,sk)（全局 L-1 坐标）。
+          const auto srcAt = [&](int si, int sj, int sk) -> std::int16_t {
+            const int cbi = si / brick - sbi0;
+            const int cbj = sj / brick - sbj0;
+            const int cbk = sk / brick - sbk0;
+            const std::vector<std::int16_t>& blk =
+                cache[(static_cast<std::size_t>(cbk) * nbj + cbj) * nbi + cbi];
+            // 块内 i 最快、k 最慢；块宽 = 该源块 extent。
+            const int lbw = extent(src.nx, si / brick, brick);
+            const int lbh = extent(src.ny, sj / brick, brick);
+            const int li = si % brick, lj = sj % brick, lk = sk % brick;
+            return blk[(static_cast<std::size_t>(lk) * lbh + lj) * lbw + li];
+          };
+
+          // 逐 dst 体素降采样（共用核）。
+          std::vector<std::int16_t> blk(static_cast<std::size_t>(bw) * bh * bd);
+          std::size_t w = 0;
+          for (int kk = 0; kk < bd; ++kk)
+            for (int jj = 0; jj < bh; ++jj)
+              for (int ii = 0; ii < bw; ++ii)
+                blk[w++] = downsampleVoxel(i0 + ii, j0 + jj, k0 + kk, src.nx,
+                                           src.ny, src.nz, srcAt);
+
+          const QByteArray compressed = compressBrick(blk);
+          const std::int64_t clen = compressed.size();
+          data.write(compressed.constData(), clen);
+          if (!data)
+            throw std::runtime_error("ChunkedVolumeStore: level data write failed");
+
+          const auto rng = computeRange(blk);
+          BrickEntry be;
+          be.offset = offset;
+          be.compressedLen = clen;
+          be.bw = bw;
+          be.bh = bh;
+          be.bd = bd;
+          be.vmin = rng.first;
+          be.vmax = rng.second;
+          be.hasRange = true;
+          lv.bricks.push_back(be);
+          offset += clen;
+        }
+      }
+    }
+    data.close();
+    rebuilt.push_back(std::move(lv));
+  }
+
+  finalizePyramidMeta(std::move(rebuilt));
+}
+
 // ----------------------- StreamingVolumeWriter -----------------------
 
 StreamingVolumeWriter::StreamingVolumeWriter(const std::string& dir,
diff --git a/src/data/store/ChunkedVolumeStore.hpp b/src/data/store/ChunkedVolumeStore.hpp
index 9cc6581..d476d81 100644
--- a/src/data/store/ChunkedVolumeStore.hpp
+++ b/src/data/store/ChunkedVolumeStore.hpp
@@ -62,6 +62,13 @@ class ChunkedVolumeStore {
   // data_L<level>.bin）。levels<=0 视为无金字塔（仅 level 0）。
   void buildPyramid(int levels);
 
+  // 流式构建金字塔：level L 的每块由 level L-1 的对应 ≤2×2×2 邻块（从盘 readBrick
+  // 读）降采样 + 逐块增量写得到，**不重组整卷**——任意时刻只持几个 L-1 邻块 + 一个
+  // L 块。产出（各级 dims/每块体素/min/max/hasRange/meta）与 buildPyramid(levels)
+  // 逐块一致。降采样/blank/min-max 规则复用 buildPyramid 同一核（DRY）。
+  // 需先有 level0 store（write 或流式建体产出）。levels<=0 视为仅 level 0。
+  void buildPyramidStreaming(int levels);
+
   // 总层数（含 level 0）；未建金字塔时 = 1。
   int levels() const { return levelCount_; }
 
@@ -108,6 +115,12 @@ class ChunkedVolumeStore {
   std::vector<std::int16_t> readBrickFrom(const Level& lv, int bx, int by,
                                           int bz) const;
 
+  // 用 rebuilt 各级（已落盘各自数据文件、每块带 offset/clen/min/max/hasRange）
+  // 替换 levels_，回写兼容字段 bricks_/levelCount_，并重写 meta.json 的 levels
+  // 数组 + level0 兼容 bricks 的 min/max。buildPyramid 与 buildPyramidStreaming
+  // 共用此收尾，确保两者 meta 结构逐字段一致（DRY）。
+  void finalizePyramidMeta(std::vector<Level> rebuilt);
+
   std::string dir_;
   StoreMeta meta_;
   int bricksX_ = 0, bricksY_ = 0, bricksZ_ = 0;  // = level 0 块数（兼容字段）
diff --git a/tests/data/store/test_pyramid.cpp b/tests/data/store/test_pyramid.cpp
index cc44613..d7a9704 100644
--- a/tests/data/store/test_pyramid.cpp
+++ b/tests/data/store/test_pyramid.cpp
@@ -144,6 +144,121 @@ TEST(Pyramid, LegacyRealAllZeroBrickRangeIsZeroZero) {
   std::filesystem::remove_all(dir);
 }
 
+// ----------------------- 流式金字塔对拍 -----------------------
+
+namespace {
+// 体素随 (i,j,k) 变化，确保逐块对拍真正区分块内容（非全相同），含 blank 散点。
+geopro::core::BuiltI16 makeVaried(int nx, int ny, int nz) {
+  geopro::core::BuiltI16 b;
+  b.vol = geopro::core::ScalarVolumeI16(nx, ny, nz);
+  for (int k = 0; k < nz; ++k)
+    for (int j = 0; j < ny; ++j)
+      for (int i = 0; i < nx; ++i) {
+        const int v = (i * 131 + j * 17 + k * 7) % 251;
+        // 散点 blank，验降采样 blank 混合（部分 blank 取非 blank 均值）。
+        b.vol.at(i, j, k) =
+            ((i + j + k) % 13 == 0)
+                ? geopro::core::ScalarVolumeI16::kBlank
+                : static_cast<std::int16_t>(v);
+      }
+  b.quant = {0.5, -3.0};
+  b.origin = {{10.0, 20.0, 30.0}};
+  b.spacing = {{2.0, 3.0, 4.0}};
+  b.vminPhys = -3.0;
+  b.vmaxPhys = 122.0;
+  return b;
+}
+
+// 流式 vs 重组整卷逐块对拍（dims + 每块体素 + min/max + hasRange + levels 数）。
+void expectStreamingMatchesInRam(const geopro::core::BuiltI16& b, int brick,
+                                 int levels) {
+  const auto dirA =
+      (std::filesystem::temp_directory_path() / "gpr_pyr_match_A").string();
+  const auto dirB =
+      (std::filesystem::temp_directory_path() / "gpr_pyr_match_B").string();
+  std::filesystem::remove_all(dirA);
+  std::filesystem::remove_all(dirB);
+
+  ChunkedVolumeStore::write(dirA, b, brick);
+  {
+    ChunkedVolumeStore a(dirA);
+    a.buildPyramid(levels);  // 金标准（重组整卷）
+  }
+  ChunkedVolumeStore::write(dirB, b, brick);
+  {
+    ChunkedVolumeStore bb(dirB);
+    bb.buildPyramidStreaming(levels);  // 流式
+  }
+
+  ChunkedVolumeStore A(dirA), B(dirB);
+  ASSERT_EQ(A.levels(), B.levels());
+  for (int L = 0; L < A.levels(); ++L) {
+    int anx, any, anz, bnx, bny, bnz;
+    A.dims(L, anx, any, anz);
+    B.dims(L, bnx, bny, bnz);
+    EXPECT_EQ(anx, bnx) << "level " << L << " nx";
+    EXPECT_EQ(any, bny) << "level " << L << " ny";
+    EXPECT_EQ(anz, bnz) << "level " << L << " nz";
+    ASSERT_EQ(A.bricksX(L), B.bricksX(L));
+    ASSERT_EQ(A.bricksY(L), B.bricksY(L));
+    ASSERT_EQ(A.bricksZ(L), B.bricksZ(L));
+    for (int bz = 0; bz < A.bricksZ(L); ++bz)
+      for (int by = 0; by < A.bricksY(L); ++by)
+        for (int bx = 0; bx < A.bricksX(L); ++bx) {
+          EXPECT_EQ(A.readBrick(L, bx, by, bz), B.readBrick(L, bx, by, bz))
+              << "voxels mismatch L=" << L << " (" << bx << "," << by << ","
+              << bz << ")";
+          EXPECT_EQ(A.brickRange(L, bx, by, bz), B.brickRange(L, bx, by, bz))
+              << "range mismatch L=" << L << " (" << bx << "," << by << ","
+              << bz << ")";
+        }
+  }
+  std::filesystem::remove_all(dirA);
+  std::filesystem::remove_all(dirB);
+}
+}  // namespace
+
+// 整除维度（128，brick64 → 2×2×2 满邻块）：流式与重组整卷逐块一致。
+TEST(Pyramid, StreamingMatchesInRam) {
+  expectStreamingMatchesInRam(makeVaried(128, 128, 128), 64, 2);
+}
+
+// 非整除/奇数维度（100、127），验边缘块 + 奇数维降采样一致。
+TEST(Pyramid, StreamingMatchesInRamNonDivisible) {
+  expectStreamingMatchesInRam(makeVaried(100, 100, 100), 64, 2);
+  expectStreamingMatchesInRam(makeVaried(127, 127, 127), 64, 3);
+}
+
+// 非立方 + 小 brick（多级、边缘块更碎）：流式与重组整卷逐块一致。
+TEST(Pyramid, StreamingMatchesInRamAnisotropicSmallBrick) {
+  expectStreamingMatchesInRam(makeVaried(70, 33, 50), 32, 3);
+}
+
+// 全 blank 体流式降采样：各级仍全 blank（min/max=kBlank），与重组整卷一致。
+TEST(Pyramid, StreamingMatchesInRamAllBlank) {
+  geopro::core::BuiltI16 b;
+  b.vol = geopro::core::ScalarVolumeI16(70, 70, 70);
+  for (auto& v : b.vol.data()) v = geopro::core::ScalarVolumeI16::kBlank;
+  b.quant = {1.0, 0.0};
+  b.origin = {{0, 0, 0}};
+  b.spacing = {{1, 1, 1}};
+  b.vminPhys = 0;
+  b.vmaxPhys = 0;
+  expectStreamingMatchesInRam(b, 32, 2);
+}
+
+// 流式不破坏 level0：兼容重载等价。
+TEST(Pyramid, StreamingLevel0ReadCompatUnchanged) {
+  auto dir =
+      (std::filesystem::temp_directory_path() / "gpr_pyr_s_compat").string();
+  std::filesystem::remove_all(dir);
+  ChunkedVolumeStore::write(dir, makeRamp(64), 32);
+  ChunkedVolumeStore s(dir);
+  s.buildPyramidStreaming(1);
+  EXPECT_EQ(s.readBrick(0, 0, 0), s.readBrick(0, 0, 0, 0));
+  std::filesystem::remove_all(dir);
+}
+
 // 老 store(未 buildPyramid):levels()==1;brickRange(0,...) 仍可惰性算。
 TEST(Pyramid, LegacyStoreNoPyramidLevelsIsOne) {
   auto dir = (std::filesystem::temp_directory_path() / "gpr_pyr_legacy").string();