diff --git a/assets/qwen/qwen-image-2.1-alpha-out2.png b/assets/qwen/qwen-image-2.1-alpha-out2.png
index b3db6c892c..8bef08a213 100644
Binary files a/assets/qwen/qwen-image-2.1-alpha-out2.png and b/assets/qwen/qwen-image-2.1-alpha-out2.png differ
diff --git a/src/core/ggml_extend.cpp b/src/core/ggml_extend.cpp
index c4c88a635b..c329a34271 100644
--- a/src/core/ggml_extend.cpp
+++ b/src/core/ggml_extend.cpp
@@ -465,7 +465,11 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
                               int d1,
                               int d2,
                               bool force_prec_f32,
-                              bool direct) {
+                              bool direct,
+                              float scale) {
+    if (scale != 1.f) {
+        x = ggml_ext_scale(ctx, x, scale);
+    }
     if (direct) {
         int64_t OC = w->ne[3] / IC;
         int64_t N  = x->ne[3] / IC;
@@ -502,6 +506,9 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
         }
     }
 
+    if (scale != 1.f) {
+        x = ggml_ext_scale(ctx, x, 1.f / scale);
+    }
     if (b != nullptr) {
         b = ggml_reshape_4d(ctx, b, 1, 1, 1, b->ne[0]);  // [OC, 1, 1, 1]
         x = ggml_add_inplace(ctx, x, b);
diff --git a/src/core/ggml_extend.h b/src/core/ggml_extend.h
index be88c58a3f..5747d625a6 100644
--- a/src/core/ggml_extend.h
+++ b/src/core/ggml_extend.h
@@ -154,7 +154,8 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
                               int d1              = 1,
                               int d2              = 1,
                               bool force_prec_f32 = false,
-                              bool direct         = false);
+                              bool direct         = false,
+                              float scale         = 1.f);
 
 // w: [OC，IC, KD, 1 * 1]
 // x: [N, IC, ID, IH*IW]
diff --git a/src/model/vae/wan_vae.hpp b/src/model/vae/wan_vae.hpp
index 73cd09537a..58cff816b5 100644
--- a/src/model/vae/wan_vae.hpp
+++ b/src/model/vae/wan_vae.hpp
@@ -24,6 +24,7 @@ namespace WAN {
         std::tuple<int, int, int> padding;
         std::tuple<int, int, int> dilation;
         bool bias;
+        float scale = 1.f;
 
         void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             auto weight = tensor_storage_map.find(prefix + "weight");
@@ -60,6 +61,10 @@ namespace WAN {
               dilation(std::move(dilation)),
               bias(bias) {}
 
+        void set_scale(float scale_value) {
+            scale = scale_value;
+        }
+
         ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* cache_x = nullptr) {
             // x: [N*IC, ID, IH, IW]
             // result: x: [N*OC, ID, IH, IW]
@@ -93,14 +98,14 @@ namespace WAN {
                 x2              = ggml_ext_conv_2d(ctx->ggml_ctx, x2, w2, b,
                                                    std::get<2>(stride), std::get<1>(stride), 0, 0,
                                                    std::get<2>(dilation), std::get<1>(dilation),
-                                                   ctx->conv2d_direct_enabled);
+                                                   ctx->conv2d_direct_enabled, false, false, scale);
                 return ggml_reshape_4d(ctx->ggml_ctx, x2, x2->ne[0], x2->ne[1], 1, out_channels);
             }
             return ggml_ext_conv_3d(ctx->ggml_ctx, ctx->backend, x, w, b, in_channels,
                                     std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
                                     0, 0, 0,
                                     std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation),
-                                    false, ctx->conv3d_direct_enabled);
+                                    false, ctx->conv3d_direct_enabled, scale);
         }
     };
 
@@ -1117,6 +1122,19 @@ namespace WAN {
             } else {
                 blocks["conv2"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(z_dim, z_dim, {1, 1, 1}));
             }
+            if (version == VERSION_QWEN_IMAGE_2_1) {
+                // Keep large VAE activations within the FP16 convolution range.
+                const float conv_scale = 1.f / 128.f;
+                std::vector<GGMLBlock*> all_blocks;
+                get_all_blocks(all_blocks);
+                for (auto block : all_blocks) {
+                    if (auto conv = dynamic_cast<Conv2d*>(block)) {
+                        conv->set_scale(conv_scale);
+                    } else if (auto conv = dynamic_cast<CausalConv3d*>(block)) {
+                        conv->set_scale(conv_scale);
+                    }
+                }
+            }
         }
 
         static ggml_tensor* patchify(ggml_context* ctx,