feat: add Ming-Image Design support (#2063)

README.md

@@ -64,6 +64,7 @@ API and command-line option may change frequently.***
     - [HiDream-O1-Image](./docs/hidream_o1_image.md)
     - [Ideogram4](./docs/ideogram4.md)
     - [LLaDA-Image](./docs/llada_image.md)
+    - [Ming-Image Design](./docs/ming_image.md)
     - [PixArt](./docs/pixart.md)
   - [Image Edit Models](./docs/edit.md)
     - [FLUX.1-Kontext-dev](./docs/kontext.md)

docs/ming_image.md

@@ -0,0 +1,24 @@
+# Ming-Image
+
+[Ming-Image](https://github.com/inclusionAI/Ming-Image) 0.1 Design uses a 6B diffusion transformer (DiT), Ling-mini-2.0 for text conditioning, and the Ming-Image VAE. Text-to-image generation with RGBA output is supported.
+
+## Download weights
+
+- Download Ming-Image 0.1 Design DiT
+    - safetensors: https://huggingface.co/Comfy-Org/Ming-Image/tree/main/diffusion_models
+- Download Ling-mini-2.0 BF16
+    - safetensors: https://huggingface.co/Comfy-Org/Ming-Image/tree/main/text_encoders
+- Download Ming-Image VAE
+    - safetensors: https://huggingface.co/Comfy-Org/Ming-Image/tree/main/vae
+- Download Ling tokenizer
+    - tokenizer.json: https://huggingface.co/inclusionAI/Ming-Image-0.1-Design/blob/main/mllm/tokenizer.json
+
+The example below uses `ming_image_0.1_design_bf16.safetensors` for the DiT. You can also use `ming_image_0.1_design_int8_convrot.safetensors` with [INT8 convrot support](int8_convrot.md). Use the BF16 text encoder.
+
+## Text-to-image
+
+Pass the Ling `tokenizer.json` with `--tokenizer` and use the matching Ming-Image VAE. Image dimensions must be multiples of 16. Save the output as PNG to preserve the alpha channel.
+
+```bash
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\ming_image_0.1_design_bf16.safetensors --llm ..\models\text_encoders\ming_image_0.1_ling_mini_2.0_bf16.safetensors --vae ..\models\vae\ming_image_vae_bf16.safetensors --tokenizer ..\models\text_encoders\tokenizer.json -p "A cheerful orange cat sticker, transparent background" --width 1024 --height 1024 --steps 12 --cfg-scale 1 --sampling-method euler --diffusion-fa -v --offload-to-cpu -o ming_image.png
+```

src/conditioning/conditioner.hpp

@@ -16,6 +16,7 @@
 #include "model/te/clip.hpp"
 #include "model/te/llada_image_te.hpp"
 #include "model/te/llm.hpp"
+#include "model/te/ming_image_te.hpp"
 #include "model/te/t5.hpp"
 #include "model_loader.h"
 #include "tokenizers/sensenova_u1_tokenizer.h"
@@ -3171,6 +3172,65 @@ struct LLMEmbedder : public Conditioner {
     }
 };
 
+struct MingImageEmbedder : public Conditioner {
+    std::shared_ptr<Tokenizer> tokenizer;
+    std::shared_ptr<MingImageTE::MingImageTextRunner> text_model;
+    const std::string prefix = "text_encoders.llm";
+
+    MingImageEmbedder(ggml_backend_t backend, const String2TensorStorage& tensors, std::shared_ptr<RunnerWeightManager> weight_manager, const TokenizerConfig& tokenizers) {
+        if (!tokenizers.has(TokenizerConfig::MAIN)) {
+            throw std::runtime_error("Ming-Image requires the Ling tokenizer.json; pass --tokenizer FILE");
+        }
+        text_model = std::make_shared<MingImageTE::MingImageTextRunner>(backend, tensors, prefix, weight_manager);
+        tokenizer  = tokenizers.create(TokenizerConfig::MAIN, text_model->config.backbone.vocab_size, 156895);
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        text_model->get_param_tensors(tensors, prefix);
+    }
+    void get_param_tensor_ops(std::map<ggml_tensor*, enum ggml_op>& ops) override {
+        text_model->get_param_tensor_ops(ops);
+    }
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        text_model->get_param_tensors(tensors, prefix);
+    }
+    void set_flash_attention_enabled(bool enabled) override { text_model->set_flash_attention_enabled(enabled); }
+    void set_max_graph_vram_bytes(size_t bytes) override { text_model->set_max_graph_vram_bytes(bytes); }
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override { text_model->set_runtime_backends(backends); }
+    void set_graph_cut_layer_split_enabled(bool enabled) override { text_model->set_graph_cut_layer_split_enabled(enabled); }
+    void set_graph_cut_layer_split_backend_vram_limits(const std::vector<size_t>& limits) override { text_model->set_graph_cut_layer_split_backend_vram_limits(limits); }
+    void set_scale_overrides(float linear, float attention) override { text_model->set_scale_overrides(linear, attention); }
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override { text_model->set_weight_adapter(adapter); }
+    void runner_end() override { text_model->runner_end(); }
+
+    SDCondition get_learned_condition(int n_threads, const ConditionerParams& input) override {
+        if (input.ref_images != nullptr && !input.ref_images->empty()) {
+            LOG_ERROR("Ming-Image currently supports text-to-image only");
+            return {};
+        }
+        std::string prompt =
+            "<role>SYSTEM</role>你是一个友好的AI助手。\n\ndetailed thinking off<|role_end|>"
+            "<role>HUMAN</role>" +
+            input.text + "<|role_end|><role>ASSISTANT</role>";
+        std::vector<int> tokens;
+        if (!tokenizer->encode(prompt, tokens, nullptr)) {
+            return {};
+        }
+        if (tokens.size() + 258 > 32768) {
+            LOG_ERROR("Ming-Image prompt exceeds the text encoder context length");
+            return {};
+        }
+        auto output = text_model->compute(n_threads, tokens);
+        if (output.empty()) {
+            return {};
+        }
+        SDCondition result;
+        result.c_crossattn = sd::ops::slice(sd::ops::slice(output, 1, 0, 256), 0, 0, text_model->config.caption_dim);
+        result.extra_c_crossattns.push_back(sd::ops::slice(output, 1, 256, output.shape()[1]));
+        return result;
+    }
+};
+
 struct LTXAVTextProjection : public GGMLBlock {
     static constexpr int64_t kHiddenSize = 3840;
     static constexpr int64_t kNumStates  = 49;

src/model.h

@@ -63,6 +63,7 @@ enum SDVersion {
     VERSION_LLADA_IMAGE,
     VERSION_ESRGAN,
     VERSION_PIXART,
+    VERSION_MING_IMAGE,
     VERSION_COUNT,
 };
 
@@ -280,7 +281,7 @@ static inline bool sd_version_uses_flux2_vae(SDVersion version) {
 }
 
 static inline bool sd_version_uses_wan_vae(SDVersion version) {
-    if (sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_qwen_image(version) || sd_version_is_krea2(version) || sd_version_is_anima(version)) {
+    if (sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_qwen_image(version) || sd_version_is_krea2(version) || sd_version_is_anima(version) || version == VERSION_MING_IMAGE) {
         return true;
     }
     return false;
@@ -314,6 +315,7 @@ static inline bool sd_version_is_dit(SDVersion version) {
         version == VERSION_HIDREAM_O1 ||
         sd_version_is_anima(version) ||
         sd_version_is_z_image(version) ||
+        version == VERSION_MING_IMAGE ||
         sd_version_is_llada_image(version) ||
         sd_version_is_boogu_image(version) ||
         sd_version_is_ernie_image(version) ||

src/model/diffusion/ming_image.hpp

@@ -0,0 +1,135 @@
+#ifndef __SD_MODEL_DIFFUSION_MING_IMAGE_HPP__
+#define __SD_MODEL_DIFFUSION_MING_IMAGE_HPP__
+
+#include "z_image.hpp"
+
+namespace MingImage {
+    struct MingImageConfig : ZImage::ZImageConfig {
+        bool split_qkv = true;
+
+        static MingImageConfig detect_from_weights(const String2TensorStorage& tensors, const std::string& prefix) {
+            MingImageConfig config;
+            static_cast<ZImage::ZImageConfig&>(config) = ZImage::ZImageConfig::detect_from_weights(tensors, prefix);
+            config.split_qkv                           = tensors.count(prefix + ".layers.0.attention.qkv.weight") == 0;
+            return config;
+        }
+    };
+
+    class MingImageModel : public GGMLBlock {
+        MingImageConfig config;
+
+    public:
+        explicit MingImageModel(const MingImageConfig& config)
+            : config(config) {
+            blocks["x_embedder"]     = std::make_shared<Linear>(config.patch_size * config.patch_size * config.in_channels, config.hidden_size);
+            blocks["t_embedder"]     = std::make_shared<TimestepEmbedder>(1024, 256, std::min<int64_t>(config.hidden_size, 256));
+            blocks["cap_embedder.0"] = std::make_shared<RMSNorm>(config.cap_feat_dim, config.norm_eps);
+            blocks["cap_embedder.1"] = std::make_shared<Linear>(config.cap_feat_dim, config.hidden_size);
+            auto add_blocks          = [&](const std::string& prefix, int64_t count, bool modulation) {
+                for (int64_t i = 0; i < count; ++i) {
+                    blocks[prefix + std::to_string(i)] = std::make_shared<ZImage::JointTransformerBlock>(
+                        static_cast<int>(i), config.hidden_size, config.head_dim, config.num_heads,
+                        config.num_kv_heads, config.multiple_of, config.ffn_dim_multiplier,
+                        config.norm_eps, config.qk_norm, modulation, true, config.split_qkv, 1e-5f);
+                }
+            };
+            add_blocks("noise_refiner.", config.num_refiner_layers, true);
+            add_blocks("context_refiner.", config.num_refiner_layers, false);
+            add_blocks("layers.", config.num_layers, true);
+            blocks["final_layer"] = std::make_shared<ZImage::FinalLayer>(config.hidden_size, config.patch_size, config.out_channels);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* timestep, ggml_tensor* context, ggml_tensor* direct, ggml_tensor* pe) {
+            auto gctx           = ctx->ggml_ctx;
+            const int64_t width = x->ne[0], height = x->ne[1];
+            auto img            = DiT::pad_and_patchify(ctx, x, config.patch_size, config.patch_size, false);
+            img                 = std::dynamic_pointer_cast<Linear>(blocks["x_embedder"])->forward(ctx, img);
+            auto txt            = std::dynamic_pointer_cast<RMSNorm>(blocks["cap_embedder.0"])->forward(ctx, context);
+            txt                 = std::dynamic_pointer_cast<Linear>(blocks["cap_embedder.1"])->forward(ctx, txt);
+            txt                 = ggml_concat(gctx, txt, direct, 1);
+            auto t              = std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder"])->forward(ctx, timestep);
+            const int64_t n_txt = txt->ne[1], n_img = img->ne[1];
+            auto txt_pe = ggml_ext_slice(gctx, pe, 3, 0, n_txt);
+            auto img_pe = ggml_ext_slice(gctx, pe, 3, n_txt, n_txt + n_img);
+            for (int64_t i = 0; i < config.num_refiner_layers; ++i) {
+                txt = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["context_refiner." + std::to_string(i)])->forward(ctx, txt, txt_pe);
+                sd::ggml_graph_cut::mark_graph_cut(txt, "ming_image.context_refiner." + std::to_string(i), "txt");
+            }
+            for (int64_t i = 0; i < config.num_refiner_layers; ++i) {
+                img = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["noise_refiner." + std::to_string(i)])->forward(ctx, img, img_pe, nullptr, t);
+                sd::ggml_graph_cut::mark_graph_cut(img, "ming_image.noise_refiner." + std::to_string(i), "img");
+            }
+            auto combined = ggml_concat(gctx, txt, img, 1);
+            for (int64_t i = 0; i < config.num_layers; ++i) {
+                combined = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["layers." + std::to_string(i)])->forward(ctx, combined, pe, nullptr, t);
+                sd::ggml_graph_cut::mark_graph_cut(combined, "ming_image.layers." + std::to_string(i), "combined");
+            }
+            img = ggml_ext_slice(gctx, combined, 1, n_txt, n_txt + n_img);
+            img = std::dynamic_pointer_cast<ZImage::FinalLayer>(blocks["final_layer"])->forward(ctx, img, t);
+            img = DiT::unpatchify_and_crop(gctx, img, height, width, config.patch_size, config.patch_size, false);
+            return ggml_scale(gctx, img, -1.f);
+        }
+    };
+
+    struct MingImageRunner : DiffusionModelRunner {
+        MingImageConfig config;
+        MingImageModel model;
+        std::vector<float> pe_values;
+
+        MingImageRunner(ggml_backend_t backend, const String2TensorStorage& tensors, const std::string& prefix, std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(MingImageConfig::detect_from_weights(tensors, prefix)),
+              model(config) {
+            model.init(params_ctx, tensors, prefix);
+        }
+
+        std::string get_desc() override { return "ming_image"; }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        sd::Tensor<float> compute(int n_threads, const DiffusionParams& inputs) override {
+            const auto* extra = diffusion_extra_as<MingImageDiffusionExtra>(inputs);
+            if (inputs.ref_latents != nullptr && !inputs.ref_latents->empty()) {
+                LOG_ERROR("Ming-Image reference-image conditioning is not supported");
+                return {};
+            }
+            if (inputs.context == nullptr || extra->direct_context == nullptr) {
+                LOG_ERROR("Ming-Image requires both query and direct text conditions");
+                return {};
+            }
+            auto graph = [&]() {
+                auto gf      = new_graph_custom(ZImage::Z_IMAGE_GRAPH_SIZE);
+                auto x       = make_input(*inputs.x);
+                auto t       = make_input(*inputs.timesteps);
+                auto context = make_input(*inputs.context);
+                auto direct  = make_input(*extra->direct_context);
+                GGML_ASSERT(x->ne[3] == 1);
+                const int64_t n_txt = context->ne[1] + direct->ne[1];
+                const int64_t n_img = ((x->ne[0] + config.patch_size - 1) / config.patch_size) *
+                                      ((x->ne[1] + config.patch_size - 1) / config.patch_size);
+                auto padded = finish_rope_pe(Rope::gen_z_image_pe(
+                    static_cast<int>(x->ne[1]), static_cast<int>(x->ne[0]), config.patch_size, 1,
+                    static_cast<int>(n_txt), ZImage::SEQ_MULTI_OF, {}, Rope::RefIndexMode::FIXED,
+                    config.theta, config.axes_dim));
+                // Zero-masked alignment tokens cannot affect valid queries. Omit them while
+                // retaining the padded caption length used to position image tokens.
+                const size_t stride      = config.axes_dim_sum * 2;
+                const int64_t padded_txt = n_txt + Rope::bound_mod(static_cast<int>(n_txt), ZImage::SEQ_MULTI_OF);
+                pe_values.assign(padded.begin(), padded.begin() + n_txt * stride);
+                pe_values.insert(pe_values.end(), padded.begin() + padded_txt * stride,
+                                 padded.begin() + (padded_txt + n_img) * stride);
+                auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, n_txt + n_img);
+                set_backend_tensor_data(pe, pe_values.data());
+                auto ctx = get_context();
+                auto out = model.forward(&ctx, x, t, context, direct, pe);
+                ggml_build_forward_expand(gf, out);
+                return gf;
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute(graph, n_threads, false), inputs.x->dim());
+        }
+    };
+}
+
+#endif  // __SD_MODEL_DIFFUSION_MING_IMAGE_HPP__

src/model/diffusion/model.hpp

@@ -141,6 +141,10 @@ struct LLaDAImageDiffusionExtra {
     const sd::Tensor<float>* semantic = nullptr;
 };
 
+struct MingImageDiffusionExtra {
+    const sd::Tensor<float>* direct_context = nullptr;
+};
+
 using DiffusionExtraParams = std::variant<std::monostate,
                                           UNetDiffusionExtra,
                                           SkipLayerDiffusionExtra,
@@ -154,7 +158,8 @@ using DiffusionExtraParams = std::variant<std::monostate,
                                           MiniT2IDiffusionExtra,
                                           SenseNovaU1DiffusionExtra,
                                           HunyuanVideoDiffusionExtra,
-                                          LLaDAImageDiffusionExtra>;
+                                          LLaDAImageDiffusionExtra,
+                                          MingImageDiffusionExtra>;
 
 struct DiffusionParams {
     const sd::Tensor<float>* x                        = nullptr;

src/model/diffusion/z_image.hpp

@@ -140,7 +140,8 @@ namespace ZImage {
                        int64_t num_kv_heads,
                        bool qk_norm,
                        bool norm_elementwise_affine = true,
-                       bool split_qkv               = false)
+                       bool split_qkv               = false,
+                       float qk_norm_eps            = 1e-6f)
             : head_dim(head_dim), num_heads(num_heads), num_kv_heads(num_kv_heads), qk_norm(qk_norm), split_qkv(split_qkv) {
             float scale = 1.f;
             if (split_qkv) {
@@ -153,8 +154,8 @@ namespace ZImage {
                 blocks["out"] = std::make_shared<Linear>(num_heads * head_dim, hidden_size, false, false, false, scale);
             }
             if (qk_norm) {
-                blocks["q_norm"] = std::make_shared<RMSNorm>(head_dim, 1e-06f, norm_elementwise_affine);
+                blocks["q_norm"] = std::make_shared<RMSNorm>(head_dim, qk_norm_eps, norm_elementwise_affine);
-                blocks["k_norm"] = std::make_shared<RMSNorm>(head_dim, 1e-06f, norm_elementwise_affine);
+                blocks["k_norm"] = std::make_shared<RMSNorm>(head_dim, qk_norm_eps, norm_elementwise_affine);
             }
         }
 
@@ -318,9 +319,10 @@ namespace ZImage {
                               bool qk_norm,
                               bool modulation              = true,
                               bool norm_elementwise_affine = true,
-                              bool split_qkv               = false)
+                              bool split_qkv               = false,
+                              float qk_norm_eps            = 1e-6f)
             : modulation(modulation) {
-            blocks["attention"]       = std::make_shared<JointAttention>(hidden_size, head_dim, num_heads, num_kv_heads, qk_norm, norm_elementwise_affine, split_qkv);
+            blocks["attention"]       = std::make_shared<JointAttention>(hidden_size, head_dim, num_heads, num_kv_heads, qk_norm, norm_elementwise_affine, split_qkv, qk_norm_eps);
             blocks["feed_forward"]    = std::make_shared<FeedForward>(hidden_size, hidden_size, multiple_of, ffn_dim_multiplier);
             blocks["attention_norm1"] = std::make_shared<RMSNorm>(hidden_size, norm_eps, norm_elementwise_affine);
             blocks["ffn_norm1"]       = std::make_shared<RMSNorm>(hidden_size, norm_eps, norm_elementwise_affine);

src/model/te/llm.hpp

@@ -50,6 +50,8 @@ namespace LLM {
         GEMMA4_12B,
         GPT_OSS_20B,
         LLADA2_MOE,
+        BAILING_MOE,
+        QWEN2,
         ARCH_COUNT,
     };
 
@@ -64,6 +66,8 @@ namespace LLM {
         "gemma4_12b",
         "gpt_oss_20b",
         "llada2_moe",
+        "bailing_moe",
+        "qwen2",
     };
 
     enum class MLPActivation {
@@ -225,7 +229,7 @@ namespace LLM {
                 config.intermediate_size       = 9216;
                 config.num_layers              = 26;
                 config.vocab_size              = 256000;
-            } else if (arch == LLMArch::LLADA2_MOE) {
+            } else if (arch == LLMArch::LLADA2_MOE || arch == LLMArch::BAILING_MOE) {
                 config.head_dim                = 128;
                 config.num_heads               = 16;
                 config.num_kv_heads            = 4;
@@ -240,7 +244,7 @@ namespace LLM {
                 config.max_position_embeddings = 16384;
                 config.rope_thetas             = {600000.f};
                 config.qkv_fused               = true;
-                config.bidirectional           = true;
+                config.bidirectional           = arch == LLMArch::LLADA2_MOE;
                 config.partial_rotary          = 0.5f;
                 config.num_experts             = 256;
                 config.num_experts_per_tok     = 8;
@@ -250,6 +254,18 @@ namespace LLM {
                 config.n_group                 = 8;
                 config.topk_group              = 4;
                 config.routed_scaling_factor   = 2.5f;
+                if (arch == LLMArch::BAILING_MOE) {
+                    config.vocab_size              = 157184;
+                    config.max_position_embeddings = 32768;
+                }
+            } else if (arch == LLMArch::QWEN2) {
+                config.hidden_size             = 1536;
+                config.intermediate_size       = 8960;
+                config.num_heads               = 12;
+                config.num_kv_heads            = 2;
+                config.vocab_size              = 151936;
+                config.max_position_embeddings = 32768;
+                config.bidirectional           = true;
             } else if (arch == LLMArch::GPT_OSS_20B) {
                 config.head_dim                = 64;
                 config.num_heads               = 64;
@@ -471,6 +487,8 @@ namespace LLM {
         int64_t n_group;
         int64_t topk_group;
         float routed_scaling_factor;
+        bool image_router;
+        bool fused_experts = false;
 
         void init_params(ggml_context* ctx,
                          const String2TensorStorage& tensor_storage_map = {},
@@ -488,6 +506,10 @@ namespace LLM {
             // scores and the group sums match.
             params["gate.weight"]      = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, hidden_size, num_experts);
             params["gate.expert_bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_experts);
+            if (image_router) {
+                params["image_gate.weight"]      = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, hidden_size, num_experts);
+                params["image_gate.expert_bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_experts);
+            }
 
             ggml_type gate_type = supported_type(get_type(prefix + "experts.gate_proj.weight", tensor_storage_map, GGML_TYPE_F32), hidden_size);
             ggml_type up_type   = supported_type(get_type(prefix + "experts.up_proj.weight", tensor_storage_map, GGML_TYPE_F32), hidden_size);
@@ -506,8 +528,14 @@ namespace LLM {
                 }
             };
 
+            fused_experts = tensor_storage_map.count(prefix + "experts.gate_up_proj.weight") != 0;
+            if (fused_experts) {
+                auto type = supported_type(get_type(prefix + "experts.gate_up_proj.weight", tensor_storage_map, GGML_TYPE_F32), hidden_size);
+                declare_experts("experts.gate_up_proj.weight", type, hidden_size, 2 * moe_intermediate_size);
+            } else {
                 declare_experts("experts.gate_proj.weight", gate_type, hidden_size, moe_intermediate_size);
                 declare_experts("experts.up_proj.weight", up_type, hidden_size, moe_intermediate_size);
+            }
             declare_experts("experts.down_proj.weight", down_type, moe_intermediate_size, hidden_size);
         }
 
@@ -519,7 +547,8 @@ namespace LLM {
               num_experts_per_tok(config.num_experts_per_tok),
               n_group(config.n_group),
               topk_group(config.topk_group),
-              routed_scaling_factor(config.routed_scaling_factor) {
+              routed_scaling_factor(config.routed_scaling_factor),
+              image_router(config.arch == LLMArch::BAILING_MOE) {
             if (config.num_shared_experts > 0) {
                 blocks["shared_experts"] = std::make_shared<MLP>(config.hidden_size,
                                                                  config.moe_intermediate_size * config.num_shared_experts,
@@ -582,7 +611,7 @@ namespace LLM {
             return ggml_mul_mat_id(ctx->ggml_ctx, w, x, selected_experts);
         }
 
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* image_mask = nullptr) {
             // x: [N, n_token, hidden_size]
             GGML_ASSERT(num_experts > 0 && num_experts_per_tok > 0);
             GGML_ASSERT(n_group > 0 && topk_group > 0 && num_experts % n_group == 0);
@@ -596,10 +625,21 @@ namespace LLM {
 
             auto logits = ggml_mul_mat(gctx, params["gate.weight"], x);
             logits      = ggml_reshape_2d(gctx, logits, num_experts, n_token_total);
-            auto scores = ggml_sigmoid(gctx, logits);  // [num_experts, tokens]
+            auto bias   = params["gate.expert_bias"];
+            if (image_router) {
+                GGML_ASSERT(image_mask != nullptr);
+                auto mask         = ggml_reshape_2d(gctx, image_mask, 1, n_token_total);
+                auto inverse_mask = ggml_scale_bias(gctx, mask, -1.f, 1.f);
+                auto image_logits = ggml_reshape_2d(gctx, ggml_mul_mat(gctx, params["image_gate.weight"], x), num_experts, n_token_total);
+                logits            = ggml_add(gctx, ggml_mul(gctx, logits, inverse_mask), ggml_mul(gctx, image_logits, mask));
+                bias              = ggml_add(gctx,
+                                             ggml_mul(gctx, ggml_repeat_4d(gctx, bias, num_experts, n_token_total, 1, 1), inverse_mask),
+                                             ggml_mul(gctx, ggml_repeat_4d(gctx, params["image_gate.expert_bias"], num_experts, n_token_total, 1, 1), mask));
+            }
+            auto scores = ggml_sigmoid(gctx, logits);
 
             // The bias steers selection only; the combine weights come from the unbiased scores.
-            auto routing = ggml_add(gctx, scores, params["gate.expert_bias"]);
+            auto routing = ggml_add(gctx, scores, bias);
             routing      = ggml_add(gctx, routing, group_limited_mask(ctx, routing, n_token_total));
 
             auto selected_experts = ggml_argsort_top_k(gctx, routing, (int)num_experts_per_tok);  // [top_k, tokens]
@@ -615,9 +655,14 @@ namespace LLM {
             weights = ggml_reshape_3d(gctx, weights, 1, num_experts_per_tok, n_token_total);
 
             auto xf = ggml_reshape_3d(gctx, x, hidden_size, 1, n_token_total);
+            ggml_tensor* activated;
+            if (fused_experts) {
+                activated = ggml_swiglu(gctx, expert_linear(ctx, "experts.gate_up_proj.weight", xf, selected_experts));
+            } else {
                 auto gate = expert_linear(ctx, "experts.gate_proj.weight", xf, selected_experts);
                 auto up   = expert_linear(ctx, "experts.up_proj.weight", xf, selected_experts);
-            auto activated = ggml_swiglu_split(gctx, gate, up);
+                activated = ggml_swiglu_split(gctx, gate, up);
+            }
             auto experts = expert_linear(ctx, "experts.down_proj.weight", activated, selected_experts);
             experts      = ggml_mul(gctx, experts, weights);
 
@@ -1428,7 +1473,9 @@ namespace LLM {
                              ggml_tensor* x,
                              ggml_tensor* input_pos,
                              ggml_tensor* attention_mask = nullptr,
-                             int rope_index              = 0) {
+                             int rope_index              = 0,
+                             ggml_tensor* rope_cos       = nullptr,
+                             ggml_tensor* rope_sin       = nullptr) {
             // x: [N, n_token, hidden_size]
             int64_t n_token = x->ne[1];
             int64_t N       = x->ne[2];
@@ -1471,13 +1518,28 @@ namespace LLM {
                 v = ggml_rms_norm(ctx->ggml_ctx, v, rms_norm_eps);
             }
 
-            if (arch == LLMArch::MISTRAL_SMALL_3_2) {
+            if (rope_cos != nullptr) {
+                GGML_ASSERT(rope_sin != nullptr);
+                // Bailing video RoPE interleaves spatial frequencies within a partial NEOX head.
+                auto rotate = [&](ggml_tensor* input) {
+                    auto gctx    = ctx->ggml_ctx;
+                    int64_t half = rope_cos->ne[0];
+                    auto first   = ggml_ext_slice(gctx, input, 0, 0, half);
+                    auto second  = ggml_ext_slice(gctx, input, 0, half, 2 * half);
+                    auto left    = ggml_sub(gctx, ggml_mul(gctx, first, rope_cos), ggml_mul(gctx, second, rope_sin));
+                    auto right   = ggml_add(gctx, ggml_mul(gctx, second, rope_cos), ggml_mul(gctx, first, rope_sin));
+                    auto rotated = ggml_concat(gctx, left, right, 0);
+                    return ggml_concat(gctx, rotated, ggml_ext_slice(gctx, input, 0, 2 * half, input->ne[0]), 0);
+                };
+                q = rotate(q);
+                k = rotate(k);
+            } else if (arch == LLMArch::MISTRAL_SMALL_3_2) {
                 q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NORMAL, 8192, 1000000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
                 k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NORMAL, 8192, 1000000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
             } else if (arch == LLMArch::MINISTRAL_3_3B) {
                 q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 262144, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
                 k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 262144, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
-            } else if (arch == LLMArch::QWEN3) {
+            } else if (arch == LLMArch::QWEN3 || arch == LLMArch::QWEN2) {
                 q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 40960, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
                 k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 40960, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
             } else if (arch == LLMArch::GPT_OSS_20B) {
@@ -1724,7 +1786,7 @@ namespace LLM {
             blocks["self_attn"] = std::make_shared<Attention>(config, sliding_attention == 0);
             if (config.arch == LLMArch::GPT_OSS_20B) {
                 blocks["mlp"] = std::make_shared<GPTOSSMLP>(config);
-            } else if (config.arch == LLMArch::LLADA2_MOE && layer_index >= config.first_k_dense_replace) {
+            } else if ((config.arch == LLMArch::LLADA2_MOE || config.arch == LLMArch::BAILING_MOE) && layer_index >= config.first_k_dense_replace) {
                 blocks["mlp"] = std::make_shared<LLaDA2MoEMLP>(config);
             } else {
                 blocks["mlp"] = std::make_shared<MLP>(config.hidden_size,
@@ -1746,7 +1808,10 @@ namespace LLM {
                              ggml_tensor* x,
                              ggml_tensor* input_pos,
                              ggml_tensor* attention_mask         = nullptr,
-                             ggml_tensor* sliding_attention_mask = nullptr) {
+                             ggml_tensor* sliding_attention_mask = nullptr,
+                             ggml_tensor* image_mask             = nullptr,
+                             ggml_tensor* rope_cos               = nullptr,
+                             ggml_tensor* rope_sin               = nullptr) {
             // x: [N, n_token, hidden_size]
             auto self_attn                                  = std::dynamic_pointer_cast<Attention>(blocks["self_attn"]);
             auto input_layernorm                            = std::dynamic_pointer_cast<LLMRMSNorm>(blocks["input_layernorm"]);
@@ -1768,7 +1833,7 @@ namespace LLM {
 
             auto residual = x;
             x             = input_layernorm->forward(ctx, x);
-            x             = self_attn->forward(ctx, x, input_pos, block_attention_mask, rope_index);
+            x             = self_attn->forward(ctx, x, input_pos, block_attention_mask, rope_index, rope_cos, rope_sin);
             if (post_attention_norm != nullptr) {
                 x = post_attention_norm->forward(ctx, x);
             }
@@ -1782,7 +1847,7 @@ namespace LLM {
             } else if (auto moe_mlp = std::dynamic_pointer_cast<LLaDA2MoEMLP>(blocks["mlp"])) {
                 // LLaDA2 is dense for the first first_k_dense_replace layers and MoE afterwards,
                 // so the block type varies per layer rather than per arch.
-                x = moe_mlp->forward(ctx, x);
+                x = moe_mlp->forward(ctx, x, image_mask);
             } else {
                 auto mlp = std::dynamic_pointer_cast<MLP>(blocks["mlp"]);
                 x        = mlp->forward(ctx, x);
@@ -1804,10 +1869,11 @@ namespace LLM {
     protected:
         int64_t num_layers;
         LLMConfig config;
+        std::string graph_cut_prefix;
 
     public:
-        TextModel(const LLMConfig& config)
+        TextModel(const LLMConfig& config, const std::string& graph_cut_prefix = "llm.text")
-            : num_layers(config.num_layers), config(config) {
+            : num_layers(config.num_layers), config(config), graph_cut_prefix(graph_cut_prefix) {
             blocks["embed_tokens"] = std::shared_ptr<GGMLBlock>(new Embedding(config.vocab_size, config.hidden_size));
             for (int i = 0; i < num_layers; i++) {
                 blocks["layers." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new TransformerBlock(config, i));
@@ -1831,7 +1897,10 @@ namespace LLM {
                                     std::set<int> out_layers,
                                     const std::vector<std::vector<std::pair<int, ggml_tensor*>>>& deepstack_image_embeds = {},
                                     ggml_tensor* sliding_attention_mask                                                  = nullptr,
-                                    bool return_all_hidden_states                                                        = false) {
+                                    bool return_all_hidden_states                                                        = false,
+                                    ggml_tensor* image_mask                                                              = nullptr,
+                                    ggml_tensor* rope_cos                                                                = nullptr,
+                                    ggml_tensor* rope_sin                                                                = nullptr) {
             auto norm = config.final_norm ? std::dynamic_pointer_cast<LLMRMSNorm>(blocks["norm"])
                                           : nullptr;
             std::vector<ggml_tensor*> intermediate_outputs;
@@ -1843,18 +1912,18 @@ namespace LLM {
                 intermediate_outputs.push_back(x);
             }
 
-            sd::ggml_graph_cut::mark_graph_cut(x, "llm.text.prelude", "x");
+            sd::ggml_graph_cut::mark_graph_cut(x, graph_cut_prefix + ".prelude", "x");
             for (int i = 0; i < num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<TransformerBlock>(blocks["layers." + std::to_string(i)]);
 
-                x = block->forward(ctx, x, input_pos, attention_mask, sliding_attention_mask);
+                x = block->forward(ctx, x, input_pos, attention_mask, sliding_attention_mask, image_mask, rope_cos, rope_sin);
                 if (i < static_cast<int>(deepstack_image_embeds.size())) {
                     x = add_deepstack_image_embeds(ctx, x, deepstack_image_embeds[static_cast<size_t>(i)]);
                 }
                 if (return_all_hidden_states || out_layers.size() > 1) {
                     x = ggml_cont(ctx->ggml_ctx, x);
                 }
-                sd::ggml_graph_cut::mark_graph_cut(x, "llm.text.layers." + std::to_string(i), "x");
+                sd::ggml_graph_cut::mark_graph_cut(x, graph_cut_prefix + ".layers." + std::to_string(i), "x");
                 if (return_all_hidden_states) {
                     if (i + 1 < num_layers) {
                         intermediate_outputs.push_back(x);

src/model/te/ming_image_te.hpp

@@ -0,0 +1,159 @@
+#ifndef __SD_MODEL_TE_MING_IMAGE_TE_HPP__
+#define __SD_MODEL_TE_MING_IMAGE_TE_HPP__
+
+#include "llm.hpp"
+
+namespace MingImageTE {
+    struct MingImageTEConfig {
+        LLM::LLMConfig backbone;
+        LLM::LLMConfig connector;
+        int64_t num_queries   = 256;
+        int64_t caption_dim   = 2560;
+        int64_t diffusion_dim = 3840;
+
+        static MingImageTEConfig detect_from_weights(const String2TensorStorage& tensors, const std::string& prefix) {
+            MingImageTEConfig config;
+            for (const auto& entry : tensors) {
+                if (starts_with(entry.first, prefix + ".backbone.") &&
+                    contains(entry.first, ".mlp.experts.") && entry.second.type == GGML_TYPE_I8) {
+                    throw std::runtime_error("Ming-Image INT8/W4A8 text encoder experts are not supported; use the BF16 text encoder");
+                }
+            }
+            bool vision           = false;
+            config.backbone       = LLM::LLMConfig::detect_from_weights(tensors, prefix + ".backbone.", LLM::LLMArch::BAILING_MOE, vision);
+            config.connector      = LLM::LLMConfig::detect_from_weights(tensors, prefix + ".connector.", LLM::LLMArch::QWEN2, vision);
+            const auto query      = tensors.find(prefix + ".query_tokens_dict.16x16");
+            const auto projection = tensors.find(prefix + ".proj_out.weight");
+            const auto direct     = tensors.find(prefix + ".proj_directvlm.1.weight");
+            if (query == tensors.end() || projection == tensors.end() || direct == tensors.end()) {
+                throw std::runtime_error("Ming-Image requires the learned queries, connector and both condition projections");
+            }
+            config.num_queries   = query->second.ne[1];
+            config.caption_dim   = projection->second.ne[1];
+            config.diffusion_dim = direct->second.ne[1];
+            if (config.num_queries != 256 || config.caption_dim != 2560 || config.diffusion_dim != 3840 ||
+                config.backbone.num_layers != 20 || config.backbone.hidden_size != 2048 ||
+                config.connector.num_layers != 28 || config.connector.hidden_size != 1536) {
+                throw std::runtime_error("unsupported Ming-Image text encoder configuration");
+            }
+            LOG_VERBOSE("ming_image_te: queries = %" PRId64 ", caption_dim = %" PRId64 ", diffusion_dim = %" PRId64,
+                        config.num_queries, config.caption_dim, config.diffusion_dim);
+            return config;
+        }
+    };
+
+    struct ConnectorModel : LLM::TextModel {
+        explicit ConnectorModel(const LLM::LLMConfig& config)
+            : LLM::TextModel(config, "ming_image.connector") {
+            blocks.erase("embed_tokens");
+        }
+    };
+
+    class MingImageTextModel : public GGMLBlock {
+        MingImageTEConfig config;
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensors = {}, const std::string prefix = "") override {
+            params["query_tokens_dict.16x16"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, config.backbone.hidden_size, config.num_queries);
+        }
+
+    public:
+        explicit MingImageTextModel(const MingImageTEConfig& config)
+            : config(config) {
+            blocks["backbone"]         = std::make_shared<LLM::TextModel>(config.backbone, "ming_image.backbone");
+            blocks["connector"]        = std::make_shared<ConnectorModel>(config.connector);
+            blocks["proj_in"]          = std::make_shared<Linear>(config.backbone.hidden_size, config.connector.hidden_size);
+            blocks["proj_out"]         = std::make_shared<Linear>(config.connector.hidden_size, config.caption_dim);
+            blocks["proj_directvlm.0"] = std::make_shared<RMSNorm>(config.backbone.hidden_size * 3, 1e-5f);
+            blocks["proj_directvlm.1"] = std::make_shared<Linear>(config.backbone.hidden_size * 3, config.diffusion_dim);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* ids, int64_t prompt_length, ggml_tensor* mask, ggml_tensor* image_mask, ggml_tensor* cos, ggml_tensor* sin, ggml_tensor* connector_positions) {
+            auto gctx                 = ctx->ggml_ctx;
+            auto backbone             = std::dynamic_pointer_cast<LLM::TextModel>(blocks["backbone"]);
+            auto connector            = std::dynamic_pointer_cast<ConnectorModel>(blocks["connector"]);
+            auto x                    = backbone->embed(ctx, ids);
+            const int64_t query_start = prompt_length + 1;
+            auto before               = ggml_ext_slice(gctx, x, 1, 0, query_start);
+            auto after                = ggml_ext_slice(gctx, x, 1, query_start + config.num_queries, x->ne[1]);
+            x                         = ggml_concat(gctx, ggml_concat(gctx, before, params["query_tokens_dict.16x16"], 1), after, 1);
+            // HF hidden_states[20] includes final RMSNorm; sd.cpp selects it as num_layers + 1.
+            x            = backbone->forward_embeds(ctx, x, nullptr, mask, {5, 12, 21}, {}, nullptr, false, image_mask, cos, sin);
+            auto direct  = ggml_ext_slice(gctx, x, 1, 0, prompt_length);
+            direct       = std::dynamic_pointer_cast<RMSNorm>(blocks["proj_directvlm.0"])->forward(ctx, direct);
+            direct       = std::dynamic_pointer_cast<Linear>(blocks["proj_directvlm.1"])->forward(ctx, direct);
+            auto queries = ggml_ext_slice(gctx, x, 0, config.backbone.hidden_size * 2, config.backbone.hidden_size * 3);
+            queries      = ggml_ext_slice(gctx, queries, 1, query_start, query_start + config.num_queries);
+            queries      = std::dynamic_pointer_cast<Linear>(blocks["proj_in"])->forward(ctx, queries);
+            queries      = connector->forward_embeds(ctx, queries, connector_positions, nullptr, {});
+            queries      = std::dynamic_pointer_cast<Linear>(blocks["proj_out"])->forward(ctx, queries);
+            queries      = ggml_pad(gctx, queries, static_cast<int>(config.diffusion_dim - config.caption_dim), 0, 0, 0);
+            return ggml_concat(gctx, queries, direct, 1);
+        }
+    };
+
+    struct MingImageTextRunner : GGMLRunner {
+        MingImageTEConfig config;
+        MingImageTextModel model;
+
+        MingImageTextRunner(ggml_backend_t backend, const String2TensorStorage& tensors, const std::string& prefix, std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : GGMLRunner(backend, weight_manager), config(MingImageTEConfig::detect_from_weights(tensors, prefix)), model(config) {
+            model.init(params_ctx, tensors, prefix);
+        }
+
+        std::string get_desc() override { return "ming_image_text"; }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        void get_param_tensor_ops(std::map<ggml_tensor*, enum ggml_op>& ops) {
+            model.get_param_tensor_ops(ops);
+        }
+
+        sd::Tensor<float> compute(int n_threads, const std::vector<int>& tokens) {
+            const int64_t prompt_length = tokens.size();
+            const int64_t query_start   = prompt_length + 1;
+            const int64_t total         = prompt_length + config.num_queries + 2;
+            std::vector<int32_t> ids(tokens.begin(), tokens.end());
+            ids.push_back(157158);
+            ids.insert(ids.end(), config.num_queries, 157157);
+            ids.push_back(157159);
+            auto input = sd::Tensor<int32_t>({total}, std::move(ids));
+            sd::Tensor<float> attention_mask({total, total});
+            sd::Tensor<float> image_mask({1, total});
+            sd::Tensor<float> cos({32, 1, total}), sin({32, 1, total});
+            std::vector<int32_t> positions(config.num_queries);
+            std::iota(positions.begin(), positions.end(), 0);
+            auto connector_positions = sd::Tensor<int32_t>({config.num_queries}, std::move(positions));
+            for (int64_t token = 0; token < total; ++token) {
+                const bool query  = token >= query_start && token < query_start + config.num_queries;
+                image_mask[token] = query ? 1.f : 0.f;
+                for (int64_t key = 0; key < total; ++key) {
+                    attention_mask[key + token * total] = key > token ? -INFINITY : 0.f;
+                }
+                // A 16x16 query bank is represented upstream as a [1, 2, 512] image grid,
+                // then spatially merged and centered to [1, 1, 256].
+                int64_t temporal = query ? query_start : (token == total - 1 ? query_start + 1 : token);
+                int64_t width    = query ? token - 127 : temporal;
+                for (int j = 0; j < 32; ++j) {
+                    int64_t position    = query && j < 24 && j % 2 ? width : temporal;
+                    float frequency     = 1.f / std::pow(600000.f, static_cast<float>(2 * j) / 64.f);
+                    float angle         = static_cast<float>(position) * frequency;
+                    cos[token * 32 + j] = std::cos(angle);
+                    sin[token * 32 + j] = std::sin(angle);
+                }
+            }
+            auto graph = [&]() {
+                auto gf  = new_graph_custom(LLM::LLM_GRAPH_SIZE);
+                auto ctx = get_context();
+                auto out = model.forward(&ctx, make_input(input), prompt_length, make_input(attention_mask),
+                                         make_input(image_mask), make_input(cos), make_input(sin), make_input(connector_positions));
+                ggml_build_forward_expand(gf, out);
+                return gf;
+            };
+            return take_or_empty(GGMLRunner::compute(graph, n_threads));
+        }
+    };
+}
+
+#endif  // __SD_MODEL_TE_MING_IMAGE_TE_HPP__

src/model/vae/wan_vae.hpp

@@ -1084,7 +1084,7 @@ namespace WAN {
 
                 _conv_num     = 34;
                 _enc_conv_num = 26;
-            } else if (version == VERSION_QWEN_IMAGE_LAYERED) {
+            } else if (version == VERSION_QWEN_IMAGE_LAYERED || version == VERSION_MING_IMAGE) {
                 input_channels = 4;
             }
 
@@ -1423,11 +1423,17 @@ namespace WAN {
         }
 
         sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+            if (version == VERSION_MING_IMAGE) {
+                return latents / 8.0064f;
+            }
             auto [mean_tensor, std_tensor] = get_latents_mean_std(latents);
             return (latents * std_tensor) / scale_factor + mean_tensor;
         }
 
         sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {
+            if (version == VERSION_MING_IMAGE) {
+                return latents * 8.0064f;
+            }
             auto [mean_tensor, std_tensor] = get_latents_mean_std(latents);
             return ((latents - mean_tensor) * scale_factor) / std_tensor;
         }

src/model_loader.cpp

@@ -524,6 +524,9 @@ SDVersion ModelLoader::get_sd_version() const {
             return VERSION_LLADA_IMAGE;
         }
         if (tensor_storage.name.find("model.diffusion_model.cap_embedder.0.weight") != std::string::npos) {
+            if (tensor_storage_map.find("text_encoders.llm.connector.layers.0.self_attn.q_proj.weight") != tensor_storage_map.end()) {
+                return VERSION_MING_IMAGE;
+            }
             return VERSION_Z_IMAGE;
         }
         if (tensor_storage.name.find("double_stream_layers.0.img_instruct_attn.processor.img_to_q.weight") != std::string::npos) {

src/name_conversion.cpp

@@ -105,7 +105,25 @@ std::string convert_open_clip_to_hf_clip_name(std::string name) {
 
 std::string convert_llada2_moe_te_name(std::string name);
 
-std::string convert_cond_stage_model_name(std::string name, std::string prefix) {
+static std::string convert_ming_image_te_name(std::string name) {
+    if (name == "llm.query_tokens") {
+        name = "llm.query_tokens_dict.16x16";
+    }
+    static const std::vector<std::pair<std::string, std::string>> name_map = {
+        {"thinker.", "backbone."},
+        {"attention.", "self_attn."},
+        {"self_attn.dense.", "self_attn.o_proj."},
+        {"gate.proj.", "gate."},
+    };
+    replace_with_name_map(name, name_map);
+    return name;
+}
+
+std::string convert_cond_stage_model_name(std::string name, std::string prefix, SDVersion version) {
+    if (version == VERSION_MING_IMAGE && prefix == "text_encoders." && starts_with(name, "llm.")) {
+        name = convert_ming_image_te_name(name);
+    }
+
     static const std::vector<std::pair<std::string, std::string>> clip_name_map{
         {"transformer.text_projection.weight", "transformer.text_model.text_projection"},
         {"model.text_projection.weight", "transformer.text_model.text_projection"},
@@ -994,6 +1012,17 @@ static std::string convert_diffusers_dit_to_original_pixart(std::string name) {
     return prefix + name;
 }
 
+static std::string convert_ming_image_dit_name(std::string name) {
+    static const std::vector<std::pair<std::string, std::string>> name_map = {
+        {"all_x_embedder.2-1.", "x_embedder."},
+        {"all_final_layer.2-1.", "final_layer."},
+        {"attention.norm_q.", "attention.q_norm."},
+        {"attention.norm_k.", "attention.k_norm."},
+    };
+    replace_with_name_map(name, name_map);
+    return name;
+}
+
 std::string convert_diffusion_model_name(std::string name, std::string prefix, SDVersion version) {
     if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
         name = convert_diffusers_unet_to_original_sd1(name);
@@ -1005,6 +1034,8 @@ std::string convert_diffusion_model_name(std::string name, std::string prefix, S
         name = convert_diffusers_dit_to_original_flux(name);
     } else if (sd_version_is_hunyuan_video(version)) {
         name = convert_hunyuan_video_to_original_flux(name);
+    } else if (version == VERSION_MING_IMAGE) {
+        name = convert_ming_image_dit_name(name);
     } else if (sd_version_is_z_image(version)) {
         name = convert_diffusers_dit_to_original_lumina2(name);
     } else if (sd_version_is_llada_image(version)) {
@@ -1668,7 +1699,7 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
     {
         for (const auto& prefix : cond_stage_model_prefix_vec) {
             if (starts_with(name, prefix)) {
-                name = convert_cond_stage_model_name(name.substr(prefix.size()), prefix);
+                name = convert_cond_stage_model_name(name.substr(prefix.size()), prefix, version);
                 name = prefix + name;
                 break;
             }

src/pipeline/diffusion_engine.cpp

@@ -106,6 +106,7 @@ const char* model_version_to_str[] = {
     "LLaDA-Image",
     "ESRGAN",
     "PixArt",
+    "Ming-Image",
 };
 
 static_assert(VERSION_COUNT == sizeof(model_version_to_str) / sizeof(model_version_to_str[0]),
@@ -1210,6 +1211,11 @@ bool StableDiffusionGGML::validate_and_load_runners() {
     LOG_VERBOSE("validating model metadata");
 
     std::set<std::string> ignore_tensors;
+    if (version == VERSION_MING_IMAGE) {
+        ignore_tensors.insert("text_encoders.llm.vision.");
+        ignore_tensors.insert("text_encoders.llm.linear_proj.");
+        ignore_tensors.insert("text_encoders.llm.backbone.lm_head.");
+    }
     if (use_tae && !tae_preview_only) {
         ignore_tensors.insert("first_stage_model.");
     }
@@ -1370,6 +1376,7 @@ bool StableDiffusionGGML::build_denoiser() {
                    sd_version_is_anima(version) ||
                    sd_version_is_ernie_image(version) ||
                    sd_version_is_z_image(version) ||
+                   version == VERSION_MING_IMAGE ||
                    sd_version_is_llada_image(version) ||
                    sd_version_is_boogu_image(version) ||
                    sd_version_is_pid(version) ||
@@ -1391,6 +1398,8 @@ bool StableDiffusionGGML::build_denoiser() {
                 default_flow_shift = 3.16f;
             } else if (sd_version_is_mage_flow(version)) {
                 default_flow_shift = 6.f;
+            } else if (version == VERSION_MING_IMAGE) {
+                default_flow_shift = INFINITY;
             } else if (sd_version_is_llada_image(version)) {
                 default_flow_shift = 1.0f;  // unused: LLADA_IMAGE_SCHEDULER builds a fixed grid
             } else {
@@ -1451,6 +1460,8 @@ bool StableDiffusionGGML::build_denoiser() {
             } else if (sd_version_is_minimax_h3(version)) {
                 LOG_INFO("running in MiniMax H3 AV FLOW mode");
                 denoiser = std::make_shared<H3AVFlowDenoiser>(default_flow_shift, 3.f, get_latent_channel());
+            } else if (version == VERSION_MING_IMAGE) {
+                denoiser = std::make_shared<MingImageFlowDenoiser>();
             } else {
                 LOG_INFO("running in FLOW mode");
                 denoiser = std::make_shared<DiscreteFlowDenoiser>();
@@ -2161,7 +2172,7 @@ std::vector<float> StableDiffusionGGML::prepare_sample_timesteps(float sigma,
     if (version == VERSION_HIDREAM_O1) {
         return std::vector<float>{1.0f - (t / static_cast<float>(TIMESTEPS))};
     }
-    if (sd_version_is_z_image(version) || sd_version_is_ideogram4(version)) {
+    if (sd_version_is_z_image(version) || sd_version_is_ideogram4(version) || version == VERSION_MING_IMAGE) {
         return std::vector<float>{1000.f - t};
     }
     return std::vector<float>{t};
@@ -2514,6 +2525,9 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
                     condition.c_token_types.empty() ? nullptr : &condition.c_token_types,
                     condition.c_vinput_mask.empty() ? nullptr : &condition.c_vinput_mask,
                     condition.c_image_embeds.empty() ? nullptr : &condition.c_image_embeds};
+            } else if (version == VERSION_MING_IMAGE) {
+                diffusion_params.extra = MingImageDiffusionExtra{
+                    condition.extra_c_crossattns.empty() ? nullptr : &condition.extra_c_crossattns[0]};
             } else if (sd_version_is_llada_image(version)) {
                 diffusion_params.extra = LLaDAImageDiffusionExtra{
                     condition.extra_c_crossattns.empty() ? nullptr : &condition.extra_c_crossattns[0]};
@@ -2748,7 +2762,7 @@ int StableDiffusionGGML::get_diffusion_model_down_factor() {
     if (sd_version_is_dit(version)) {
         if (sd_version_is_sensenova_u1(version)) {
             down_factor = 32;
-        } else if (version == VERSION_QWEN_IMAGE_2_1 || sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_minimax_h3(version) || sd_version_is_pixart(version)) {
+        } else if (version == VERSION_QWEN_IMAGE_2_1 || version == VERSION_MING_IMAGE || sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_minimax_h3(version) || sd_version_is_pixart(version)) {
             down_factor = 2;
         } else {
             down_factor = 1;
@@ -2796,7 +2810,7 @@ int StableDiffusionGGML::get_latent_channel() {
 }
 
 int StableDiffusionGGML::get_image_channels() const {
-    return version == VERSION_QWEN_IMAGE_LAYERED || version == VERSION_QWEN_IMAGE_2_1 ? 4 : 3;
+    return version == VERSION_QWEN_IMAGE_LAYERED || version == VERSION_QWEN_IMAGE_2_1 || version == VERSION_MING_IMAGE ? 4 : 3;
 }
 
 int StableDiffusionGGML::get_image_seq_len(int h, int w) {

src/pipeline/image.cpp

@@ -465,6 +465,11 @@ namespace sd::pipeline {
                 // states with a zeroed prompt mask, so no extra text encode is needed.
                 uncond.c_crossattn = cond.c_crossattn;
                 uncond.c_vector    = sd::Tensor<float>::zeros_like(cond.c_vector);
+            } else if (sd->version == VERSION_MING_IMAGE) {
+                uncond.c_crossattn = sd::Tensor<float>::zeros_like(cond.c_crossattn);
+                for (const auto& extra : cond.extra_c_crossattns) {
+                    uncond.extra_c_crossattns.push_back(sd::Tensor<float>::zeros_like(extra));
+                }
             } else if (sd_version_is_sensenova_u1(sd->version)) {
                 auto* sensenova_conditioner = static_cast<SenseNovaU1Conditioner*>(sd->cond_stage_model.get());
                 uncond                      = sensenova_conditioner->get_unconditional_condition(request->negative_prompt);

src/pipeline/model_builders.cpp

@@ -24,6 +24,7 @@
 #include "model/diffusion/llada_image.hpp"
 #include "model/diffusion/ltxv.hpp"
 #include "model/diffusion/mage_flow.hpp"
+#include "model/diffusion/ming_image.hpp"
 #include "model/diffusion/minimax_h3.hpp"
 #include "model/diffusion/minit2i.hpp"
 #include "model/diffusion/mmdit.hpp"
@@ -374,6 +375,11 @@ namespace sd::model_builders {
                                                                     tensor_storage_map,
                                                                     "model.diffusion_model",
                                                                     weight_manager);
+        } else if (version == VERSION_MING_IMAGE) {
+            result.conditioner = std::make_shared<MingImageEmbedder>(ctx.backends.runtime_backend(SDBackendModule::TE),
+                                                                     tensor_storage_map, weight_manager, tokenizers);
+            result.diffusion   = std::make_shared<MingImage::MingImageRunner>(ctx.backends.runtime_backend(SDBackendModule::DIFFUSION),
+                                                                            tensor_storage_map, "model.diffusion_model", weight_manager);
         } else if (sd_version_is_z_image(version)) {
             result.conditioner = std::make_shared<LLMEmbedder>(ctx.backends.runtime_backend(SDBackendModule::TE),
                                                                tensor_storage_map,

src/runtime/denoiser.hpp

@@ -1369,6 +1369,42 @@ struct DiscreteFlowDenoiser : public Denoiser {
     }
 };
 
+struct MingImageFlowDenoiser : DiscreteFlowDenoiser {
+    float resolution_shift = std::exp(1.35f);
+
+    MingImageFlowDenoiser()
+        : DiscreteFlowDenoiser(INFINITY) {}
+
+    float sigma_min() override { return 0.f; }
+    float sigma_max() override { return 1.f; }
+
+    float t_to_sigma(float t) override {
+        return time_snr_shift(std::isfinite(shift) ? shift : resolution_shift, (t + 1.f) / 1000.f);
+    }
+
+    std::vector<float> get_sigmas(uint32_t n, int image_seq_len, scheduler_t scheduler, SDVersion version, const char* extra_sample_args = nullptr) override {
+        const float tokens = static_cast<float>(image_seq_len) / 4.f;
+        const float mu     = tokens >= 4096.f ? 1.35f : 0.5f + (tokens - 256.f) * (1.15f - 0.5f) / (4096.f - 256.f);
+        resolution_shift   = std::exp(mu);
+        if (scheduler != DISCRETE_SCHEDULER) {
+            return DiscreteFlowDenoiser::get_sigmas(n, image_seq_len, scheduler, version, extra_sample_args);
+        }
+        if (n == 0) {
+            return {};
+        }
+        const float factor = std::isfinite(shift) ? shift : resolution_shift;
+        std::vector<float> sigmas;
+        sigmas.reserve(n + 1);
+        for (uint32_t i = 0; i < n; ++i) {
+            const float t = n == 1 ? 1.f : 1.f - static_cast<float>(i) / static_cast<float>(n - 1);
+            sigmas.push_back(time_snr_shift(factor, t));
+        }
+        // Upstream sets sigma_min to zero before linspace, then appends the terminal zero.
+        sigmas.push_back(0.f);
+        return sigmas;
+    }
+};
+
 struct H3AVFlowDenoiser : public DiscreteFlowDenoiser {
     int64_t video_channels;
     float audio_shift;
@@ -1784,6 +1820,9 @@ static sd::Tensor<float> sample_euler(denoise_cb_t model,
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
         float sigma = sigmas[i];
+        if (sigma == sigmas[i + 1]) {
+            continue;
+        }
         auto denoised_opt = model(x, sigma, i + 1);
         if (denoised_opt.pred.empty()) {
             return {};