fix: set VAE conv scale for all SDXL variants

* feat: add basic sdapi support to sd-server

Compatible with AUTOMATIC1111 / Forge.

* fix img2img with no mask

* add more parameter validation

* eliminate MSVC warnings

---------

Co-authored-by: leejet <leejet714@gmail.com>

README.md

@@ -43,8 +43,8 @@ API and command-line option may change frequently.***
     - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
     - [Some SD1.x and SDXL distilled models](./docs/distilled_sd.md)
     - [SD3/SD3.5](./docs/sd3.md)
-    - [FlUX.1-dev/FlUX.1-schnell](./docs/flux.md)
-    - [FLUX.2-dev](./docs/flux2.md)
+    - [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
+    - [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
     - [Chroma](./docs/chroma.md)
     - [Chroma1-Radiance](./docs/chroma_radiance.md)
     - [Qwen Image](./docs/qwen_image.md)
@@ -127,8 +127,8 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 
 - [SD1.x/SD2.x/SDXL](./docs/sd.md)
 - [SD3/SD3.5](./docs/sd3.md)
-- [FlUX.1-dev/FlUX.1-schnell](./docs/flux.md)
-- [FLUX.2-dev](./docs/flux2.md)
+- [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
+- [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
 - [FLUX.1-Kontext-dev](./docs/kontext.md)
 - [Chroma](./docs/chroma.md)
 - [🔥Qwen Image](./docs/qwen_image.md)

conditioner.hpp

@@ -1614,9 +1614,9 @@ struct LLMEmbedder : public Conditioner {
                 bool enable_vision                             = false)
         : version(version) {
         LLM::LLMArch arch = LLM::LLMArch::QWEN2_5_VL;
-        if (sd_version_is_flux2(version)) {
+        if (version == VERSION_FLUX2) {
             arch = LLM::LLMArch::MISTRAL_SMALL_3_2;
-        } else if (sd_version_is_z_image(version) || version == VERSION_OVIS_IMAGE) {
+        } else if (sd_version_is_z_image(version) || version == VERSION_OVIS_IMAGE || version == VERSION_FLUX2_KLEIN) {
             arch = LLM::LLMArch::QWEN3;
         }
         if (arch == LLM::LLMArch::MISTRAL_SMALL_3_2) {
@@ -1708,6 +1708,9 @@ struct LLMEmbedder : public Conditioner {
         int prompt_template_encode_start_idx = 34;
         int max_length                       = 0;
         std::set<int> out_layers;
+        std::vector<int> tokens;
+        std::vector<float> weights;
+        std::vector<float> mask;
         if (llm->enable_vision && conditioner_params.ref_images.size() > 0) {
             LOG_INFO("QwenImageEditPlusPipeline");
             prompt_template_encode_start_idx = 64;
@@ -1771,7 +1774,7 @@ struct LLMEmbedder : public Conditioner {
             prompt_attn_range.second = static_cast<int>(prompt.size());
 
             prompt += "<|im_end|>\n<|im_start|>assistant\n";
-        } else if (sd_version_is_flux2(version)) {
+        } else if (version == VERSION_FLUX2) {
             prompt_template_encode_start_idx = 0;
             out_layers                       = {10, 20, 30};
 
@@ -1793,17 +1796,28 @@ struct LLMEmbedder : public Conditioner {
             prompt_attn_range.second = static_cast<int>(prompt.size());
 
             prompt += "<|im_end|>\n<|im_start|>assistant\n";
-        } else if (sd_version_is_flux2(version)) {
+        } else if (version == VERSION_FLUX2_KLEIN) {
             prompt_template_encode_start_idx = 0;
-            out_layers                       = {10, 20, 30};
+            max_length                       = 512;
+            out_layers                       = {9, 18, 27};
 
-            prompt = "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]";
+            prompt = "<|im_start|>user\n";
 
             prompt_attn_range.first = static_cast<int>(prompt.size());
             prompt += conditioner_params.text;
             prompt_attn_range.second = static_cast<int>(prompt.size());
 
-            prompt += "[/INST]";
+            prompt += "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n";
+
+            auto tokens_and_weights = tokenize(prompt, prompt_attn_range, 0, false);
+            tokens                  = std::get<0>(tokens_and_weights);
+            weights                 = std::get<1>(tokens_and_weights);
+
+            mask.insert(mask.end(), tokens.size(), 1.f);
+            if (tokens.size() < max_length) {
+                mask.insert(mask.end(), max_length - tokens.size(), 0.f);
+                tokenizer->pad_tokens(tokens, weights, max_length, true);
+            }
         } else if (version == VERSION_OVIS_IMAGE) {
             prompt_template_encode_start_idx = 28;
             max_length                       = prompt_template_encode_start_idx + 256;
@@ -1827,17 +1841,34 @@ struct LLMEmbedder : public Conditioner {
             prompt += "<|im_end|>\n<|im_start|>assistant\n";
         }
 
-        auto tokens_and_weights = tokenize(prompt, prompt_attn_range, max_length, max_length > 0);
-        auto& tokens            = std::get<0>(tokens_and_weights);
-        auto& weights           = std::get<1>(tokens_and_weights);
+        if (tokens.empty()) {
+            auto tokens_and_weights = tokenize(prompt, prompt_attn_range, max_length, max_length > 0);
+            tokens                  = std::get<0>(tokens_and_weights);
+            weights                 = std::get<1>(tokens_and_weights);
+        }
 
         int64_t t0                        = ggml_time_ms();
         struct ggml_tensor* hidden_states = nullptr;  // [N, n_token, 3584]
 
         auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens);
 
+        ggml_tensor* attention_mask = nullptr;
+        if (!mask.empty()) {
+            attention_mask = ggml_new_tensor_2d(work_ctx, GGML_TYPE_F32, mask.size(), mask.size());
+            ggml_ext_tensor_iter(attention_mask, [&](ggml_tensor* attention_mask, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
+                float value = 0.f;
+                if (mask[i0] == 0.f) {
+                    value = -INFINITY;
+                } else if (i0 > i1) {
+                    value = -INFINITY;
+                }
+                ggml_ext_tensor_set_f32(attention_mask, value, i0, i1, i2, i3);
+            });
+        }
+
         llm->compute(n_threads,
                      input_ids,
+                     attention_mask,
                      image_embeds,
                      out_layers,
                      &hidden_states,
@@ -1861,7 +1892,7 @@ struct LLMEmbedder : public Conditioner {
         GGML_ASSERT(hidden_states->ne[1] > prompt_template_encode_start_idx);
 
         int64_t min_length = 0;
-        if (sd_version_is_flux2(version)) {
+        if (version == VERSION_FLUX2) {
             min_length = 512;
         }
 

docs/distilled_sd.md

@@ -1,8 +1,8 @@
-# Running distilled models: SSD1B and SDx.x with tiny U-Nets
+# Running distilled models: SSD1B, Vega and SDx.x with tiny U-Nets
 
 ## Preface 
 
-These models feature a reduced U-Net architecture. Unlike standard SDXL models, the SSD-1B U-Net contains only one middle block and fewer attention layers in its up- and down-blocks, resulting in significantly smaller file sizes. Using these models can reduce inference time by more than 33%. For more details, refer to Segmind's paper: https://arxiv.org/abs/2401.02677v1.
+These models feature a reduced U-Net architecture. Unlike standard SDXL models, the SSD-1B and Vega U-Net contains only one middle block and fewer attention layers in its up- and down-blocks, resulting in significantly smaller file sizes. Using these models can reduce inference time by more than 33%. For more details, refer to Segmind's paper: https://arxiv.org/abs/2401.02677v1.
 Similarly, SD1.x- and SD2.x-style models with a tiny U-Net consist of only 6 U-Net blocks, leading to very small files and time savings of up to 50%. For more information, see the paper: https://arxiv.org/pdf/2305.15798.pdf.
 
 ## SSD1B
@@ -17,7 +17,17 @@ Useful LoRAs are also available:
  * https://huggingface.co/seungminh/lora-swarovski-SSD-1B/resolve/main/pytorch_lora_weights.safetensors
  * https://huggingface.co/kylielee505/mylcmlorassd/resolve/main/pytorch_lora_weights.safetensors
 
-These files can be used out-of-the-box, unlike the models described in the next section.
+## Vega
+
+Segmind's Vega model is available online here:
+
+ * https://huggingface.co/segmind/Segmind-Vega/resolve/main/segmind-vega.safetensors
+ 
+VegaRT is an example for an LCM-LoRA:
+
+ * https://huggingface.co/segmind/Segmind-VegaRT/resolve/main/pytorch_lora_weights.safetensors
+
+Both files can be used out-of-the-box, unlike the models described in next sections.
 
 
 ## SD1.x, SD2.x with tiny U-Nets

docs/esrgan.md

@@ -1,6 +1,6 @@
 ## Using ESRGAN to upscale results
 
-You can use ESRGAN to upscale the generated images. At the moment, only the [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth) model is supported. Support for more models of this architecture will be added soon.
+You can use ESRGAN—such as the model [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth)—to upscale the generated images and improve their overall resolution and clarity.
 
 - Specify the model path using the `--upscale-model PATH` parameter. example:
 

docs/flux2.md

@@ -1,6 +1,8 @@
 # How to Use
 
-## Download weights
+## Flux.2-dev
+
+### Download weights
 
 - Download FLUX.2-dev
     - gguf: https://huggingface.co/city96/FLUX.2-dev-gguf/tree/main
@@ -9,7 +11,7 @@
 - Download Mistral-Small-3.2-24B-Instruct-2506-GGUF
     - gguf: https://huggingface.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main
 
-## Examples
+### Examples
 
 ```
 .\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
@@ -17,5 +19,74 @@
 
 <img alt="flux2 example" src="../assets/flux2/example.png" />
 
+## Flux.2 klein 4B / Flux.2 klein base 4B
 
+### Download weights
 
+- Download FLUX.2-klein-4B
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-4B
+    - gguf: https://huggingface.co/leejet/FLUX.2-klein-4B-GGUF/tree/main
+- Download FLUX.2-klein-base-4B
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B
+    - gguf: https://huggingface.co/leejet/FLUX.2-klein-base-4B-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download Qwen3 4b
+    - safetensors: https://huggingface.co/Comfy-Org/flux2-klein-4B/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/unsloth/Qwen3-4B-GGUF/tree/main
+
+### Examples
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="flux2-klein-4b" src="../assets/flux2/flux2-klein-4b.png" />
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
+```
+
+<img alt="flux2-klein-4b-edit" src="../assets/flux2/flux2-klein-4b-edit.png" />
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="flux2-klein-base-4b" src="../assets/flux2/flux2-klein-base-4b.png" />
+
+## Flux.2 klein 9B / Flux.2 klein base 9B
+
+### Download weights
+
+- Download FLUX.2-klein-9B
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-9B
+    - gguf: https://huggingface.co/leejet/FLUX.2-klein-9B-GGUF/tree/main
+- Download FLUX.2-klein-base-9B
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B
+    - gguf: https://huggingface.co/leejet/FLUX.2-klein-base-9B-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download Qwen3 8B
+    - safetensors: https://huggingface.co/Comfy-Org/flux2-klein-9B/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/unsloth/Qwen3-8B-GGUF/tree/main
+
+### Examples
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="flux2-klein-9b" src="../assets/flux2/flux2-klein-9b.png" />
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
+```
+
+<img alt="flux2-klein-9b-edit" src="../assets/flux2/flux2-klein-9b-edit.png" />
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="flux2-klein-base-9b" src="../assets/flux2/flux2-klein-base-9b.png" />

examples/server/main.cpp

@@ -86,21 +86,6 @@ std::vector<uint8_t> base64_decode(const std::string& encoded_string) {
     return ret;
 }
 
-std::string iso_timestamp_now() {
-    using namespace std::chrono;
-    auto now      = system_clock::now();
-    std::time_t t = system_clock::to_time_t(now);
-    std::tm tm{};
-#ifdef _MSC_VER
-    gmtime_s(&tm, &t);
-#else
-    gmtime_r(&t, &tm);
-#endif
-    std::ostringstream oss;
-    oss << std::put_time(&tm, "%Y-%m-%dT%H:%M:%SZ");
-    return oss.str();
-}
-
 struct SDSvrParams {
     std::string listen_ip = "127.0.0.1";
     int listen_port       = 1234;
@@ -404,7 +389,7 @@ int main(int argc, const char** argv) {
             }
 
             json out;
-            out["created"]       = iso_timestamp_now();
+            out["created"]       = static_cast<long long>(std::time(nullptr));
             out["data"]          = json::array();
             out["output_format"] = output_format;
 
@@ -692,7 +677,7 @@ int main(int argc, const char** argv) {
             }
 
             json out;
-            out["created"]       = iso_timestamp_now();
+            out["created"]       = static_cast<long long>(std::time(nullptr));
             out["data"]          = json::array();
             out["output_format"] = output_format;
 
@@ -732,6 +717,327 @@ int main(int argc, const char** argv) {
         }
     });
 
+    // sdapi endpoints (AUTOMATIC1111 / Forge)
+
+    auto sdapi_any2img = [&](const httplib::Request& req, httplib::Response& res, bool img2img) {
+        try {
+            if (req.body.empty()) {
+                res.status = 400;
+                res.set_content(R"({"error":"empty body"})", "application/json");
+                return;
+            }
+
+            json j = json::parse(req.body);
+
+            std::string prompt          = j.value("prompt", "");
+            std::string negative_prompt = j.value("negative_prompt", "");
+            int width                   = j.value("width", 512);
+            int height                  = j.value("height", 512);
+            int steps                   = j.value("steps", -1);
+            float cfg_scale             = j.value("cfg_scale", 7.f);
+            int64_t seed                = j.value("seed", -1);
+            int batch_size              = j.value("batch_size", 1);
+            int clip_skip               = j.value("clip_skip", -1);
+            std::string sampler_name    = j.value("sampler_name", "");
+            std::string scheduler_name  = j.value("scheduler", "");
+
+            auto bad = [&](const std::string& msg) {
+                res.status = 400;
+                res.set_content("{\"error\":\"" + msg + "\"}", "application/json");
+                return;
+            };
+
+            if (width <= 0 || height <= 0) {
+                return bad("width and height must be positive");
+            }
+
+            if (steps < 1 || steps > 150) {
+                return bad("steps must be in range [1, 150]");
+            }
+
+            if (batch_size < 1 || batch_size > 8) {
+                return bad("batch_size must be in range [1, 8]");
+            }
+
+            if (cfg_scale < 0.f) {
+                return bad("cfg_scale must be positive");
+            }
+
+            if (prompt.empty()) {
+                return bad("prompt required");
+            }
+
+            auto get_sample_method = [](std::string name) -> enum sample_method_t {
+                enum sample_method_t result = str_to_sample_method(name.c_str());
+                if (result != SAMPLE_METHOD_COUNT) return result;
+                // some applications use a hardcoded sampler list
+                std::transform(name.begin(), name.end(), name.begin(),
+                               [](unsigned char c) { return std::tolower(c); });
+                static const std::unordered_map<std::string_view, sample_method_t> hardcoded{
+                    {"euler a", EULER_A_SAMPLE_METHOD},
+                    {"k_euler_a", EULER_A_SAMPLE_METHOD},
+                    {"euler", EULER_SAMPLE_METHOD},
+                    {"k_euler", EULER_SAMPLE_METHOD},
+                    {"heun", HEUN_SAMPLE_METHOD},
+                    {"k_heun", HEUN_SAMPLE_METHOD},
+                    {"dpm2", DPM2_SAMPLE_METHOD},
+                    {"k_dpm_2", DPM2_SAMPLE_METHOD},
+                    {"lcm", LCM_SAMPLE_METHOD},
+                    {"ddim", DDIM_TRAILING_SAMPLE_METHOD},
+                    {"dpm++ 2m", DPMPP2M_SAMPLE_METHOD},
+                    {"k_dpmpp_2m", DPMPP2M_SAMPLE_METHOD}};
+                auto it            = hardcoded.find(name);
+                if (it != hardcoded.end()) return it->second;
+                return SAMPLE_METHOD_COUNT;
+            };
+
+            enum sample_method_t sample_method = get_sample_method(sampler_name);
+
+            enum scheduler_t scheduler = str_to_scheduler(scheduler_name.c_str());
+
+            // avoid excessive resource usage
+
+            SDGenerationParams gen_params         = default_gen_params;
+            gen_params.prompt                     = prompt;
+            gen_params.negative_prompt            = negative_prompt;
+            gen_params.width                      = width;
+            gen_params.height                     = height;
+            gen_params.seed                       = seed;
+            gen_params.sample_params.sample_steps = steps;
+            gen_params.batch_count                = batch_size;
+
+            if (clip_skip > 0) {
+                gen_params.clip_skip = clip_skip;
+            }
+
+            if (sample_method != SAMPLE_METHOD_COUNT) {
+                gen_params.sample_params.sample_method = sample_method;
+            }
+
+            if (scheduler != SCHEDULER_COUNT) {
+                gen_params.sample_params.scheduler = scheduler;
+            }
+
+            LOG_DEBUG("%s\n", gen_params.to_string().c_str());
+
+            sd_image_t init_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+            sd_image_t control_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+            sd_image_t mask_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 1, nullptr};
+            std::vector<uint8_t> mask_data;
+            std::vector<sd_image_t> pmid_images;
+            std::vector<sd_image_t> ref_images;
+
+            if (img2img) {
+                auto decode_image = [](sd_image_t& image, std::string encoded) -> bool {
+                    // remove data URI prefix if present ("data:image/png;base64,")
+                    auto comma_pos = encoded.find(',');
+                    if (comma_pos != std::string::npos) {
+                        encoded = encoded.substr(comma_pos + 1);
+                    }
+                    std::vector<uint8_t> img_data = base64_decode(encoded);
+                    if (!img_data.empty()) {
+                        int img_w         = image.width;
+                        int img_h         = image.height;
+                        uint8_t* raw_data = load_image_from_memory(
+                            (const char*)img_data.data(), (int)img_data.size(),
+                            img_w, img_h,
+                            image.width, image.height, image.channel);
+                        if (raw_data) {
+                            image = {(uint32_t)img_w, (uint32_t)img_h, image.channel, raw_data};
+                            return true;
+                        }
+                    }
+                    return false;
+                };
+
+                if (j.contains("init_images") && j["init_images"].is_array() && !j["init_images"].empty()) {
+                    std::string encoded = j["init_images"][0].get<std::string>();
+                    decode_image(init_image, encoded);
+                }
+
+                if (j.contains("mask") && j["mask"].is_string()) {
+                    std::string encoded = j["mask"].get<std::string>();
+                    decode_image(mask_image, encoded);
+                    bool inpainting_mask_invert = j.value("inpainting_mask_invert", 0) != 0;
+                    if (inpainting_mask_invert && mask_image.data != nullptr) {
+                        for (uint32_t i = 0; i < mask_image.width * mask_image.height; i++) {
+                            mask_image.data[i] = 255 - mask_image.data[i];
+                        }
+                    }
+                } else {
+                    mask_data          = std::vector<uint8_t>(width * height, 255);
+                    mask_image.width   = width;
+                    mask_image.height  = height;
+                    mask_image.channel = 1;
+                    mask_image.data    = mask_data.data();
+                }
+
+                if (j.contains("extra_images") && j["extra_images"].is_array()) {
+                    for (auto extra_image : j["extra_images"]) {
+                        std::string encoded  = extra_image.get<std::string>();
+                        sd_image_t tmp_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+                        if (decode_image(tmp_image, encoded)) {
+                            ref_images.push_back(tmp_image);
+                        }
+                    }
+                }
+
+                float denoising_strength = j.value("denoising_strength", -1.f);
+                if (denoising_strength >= 0.f) {
+                    denoising_strength  = std::min(denoising_strength, 1.0f);
+                    gen_params.strength = denoising_strength;
+                }
+            }
+
+            sd_img_gen_params_t img_gen_params = {
+                gen_params.lora_vec.data(),
+                static_cast<uint32_t>(gen_params.lora_vec.size()),
+                gen_params.prompt.c_str(),
+                gen_params.negative_prompt.c_str(),
+                gen_params.clip_skip,
+                init_image,
+                ref_images.data(),
+                (int)ref_images.size(),
+                gen_params.auto_resize_ref_image,
+                gen_params.increase_ref_index,
+                mask_image,
+                gen_params.width,
+                gen_params.height,
+                gen_params.sample_params,
+                gen_params.strength,
+                gen_params.seed,
+                gen_params.batch_count,
+                control_image,
+                gen_params.control_strength,
+                {
+                    pmid_images.data(),
+                    (int)pmid_images.size(),
+                    gen_params.pm_id_embed_path.c_str(),
+                    gen_params.pm_style_strength,
+                },  // pm_params
+                ctx_params.vae_tiling_params,
+                gen_params.cache_params,
+            };
+
+            sd_image_t* results = nullptr;
+            int num_results     = 0;
+
+            {
+                std::lock_guard<std::mutex> lock(sd_ctx_mutex);
+                results     = generate_image(sd_ctx, &img_gen_params);
+                num_results = gen_params.batch_count;
+            }
+
+            json out;
+            out["images"]     = json::array();
+            out["parameters"] = j;  // TODO should return changed defaults
+            out["info"]       = "";
+
+            for (int i = 0; i < num_results; i++) {
+                if (results[i].data == nullptr) {
+                    continue;
+                }
+
+                auto image_bytes = write_image_to_vector(ImageFormat::PNG,
+                                                         results[i].data,
+                                                         results[i].width,
+                                                         results[i].height,
+                                                         results[i].channel);
+
+                if (image_bytes.empty()) {
+                    LOG_ERROR("write image to mem failed");
+                    continue;
+                }
+
+                std::string b64 = base64_encode(image_bytes);
+                out["images"].push_back(b64);
+            }
+
+            res.set_content(out.dump(), "application/json");
+            res.status = 200;
+
+            if (init_image.data) {
+                stbi_image_free(init_image.data);
+            }
+            if (mask_image.data && mask_data.empty()) {
+                stbi_image_free(mask_image.data);
+            }
+            for (auto ref_image : ref_images) {
+                stbi_image_free(ref_image.data);
+            }
+
+        } catch (const std::exception& e) {
+            res.status = 500;
+            json err;
+            err["error"]   = "server_error";
+            err["message"] = e.what();
+            res.set_content(err.dump(), "application/json");
+        }
+    };
+
+    svr.Post("/sdapi/v1/txt2img", [&](const httplib::Request& req, httplib::Response& res) {
+        sdapi_any2img(req, res, false);
+    });
+
+    svr.Post("/sdapi/v1/img2img", [&](const httplib::Request& req, httplib::Response& res) {
+        sdapi_any2img(req, res, true);
+    });
+
+    svr.Get("/sdapi/v1/samplers", [&](const httplib::Request&, httplib::Response& res) {
+        std::vector<std::string> sampler_names;
+        sampler_names.push_back("default");
+        for (int i = 0; i < SAMPLE_METHOD_COUNT; i++) {
+            sampler_names.push_back(sd_sample_method_name((sample_method_t)i));
+        }
+        json r = json::array();
+        for (auto name : sampler_names) {
+            json entry;
+            entry["name"]    = name;
+            entry["aliases"] = json::array({name});
+            entry["options"] = json::object();
+            r.push_back(entry);
+        }
+        res.set_content(r.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/schedulers", [&](const httplib::Request&, httplib::Response& res) {
+        std::vector<std::string> scheduler_names;
+        scheduler_names.push_back("default");
+        for (int i = 0; i < SCHEDULER_COUNT; i++) {
+            scheduler_names.push_back(sd_scheduler_name((scheduler_t)i));
+        }
+        json r = json::array();
+        for (auto name : scheduler_names) {
+            json entry;
+            entry["name"]  = name;
+            entry["label"] = name;
+            r.push_back(entry);
+        }
+        res.set_content(r.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/sd-models", [&](const httplib::Request&, httplib::Response& res) {
+        fs::path model_path = ctx_params.model_path;
+        json entry;
+        entry["title"]      = model_path.stem();
+        entry["model_name"] = model_path.stem();
+        entry["filename"]   = model_path.filename();
+        entry["hash"]       = "8888888888";
+        entry["sha256"]     = "8888888888888888888888888888888888888888888888888888888888888888";
+        entry["config"]     = nullptr;
+        json r              = json::array();
+        r.push_back(entry);
+        res.set_content(r.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/options", [&](const httplib::Request&, httplib::Response& res) {
+        fs::path model_path = ctx_params.model_path;
+        json r;
+        r["samples_format"]      = "png";
+        r["sd_model_checkpoint"] = model_path.stem();
+        res.set_content(r.dump(), "application/json");
+    });
+
     LOG_INFO("listening on: %s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
     svr.listen(svr_params.listen_ip, svr_params.listen_port);
 

flux.hpp

@@ -1288,13 +1288,9 @@ namespace Flux {
             } else if (version == VERSION_OVIS_IMAGE) {
                 flux_params.semantic_txt_norm = true;
                 flux_params.use_yak_mlp       = true;
-                flux_params.context_in_dim    = 2048;
                 flux_params.vec_in_dim        = 0;
             } else if (sd_version_is_flux2(version)) {
-                flux_params.context_in_dim   = 15360;
                 flux_params.in_channels      = 128;
-                flux_params.hidden_size      = 6144;
-                flux_params.num_heads        = 48;
                 flux_params.patch_size       = 1;
                 flux_params.out_channels     = 128;
                 flux_params.mlp_ratio        = 3.f;
@@ -1307,12 +1303,12 @@ namespace Flux {
                 flux_params.ref_index_scale  = 10.f;
                 flux_params.use_mlp_silu_act = true;
             }
+            int64_t head_dim = 0;
             for (auto pair : tensor_storage_map) {
                 std::string tensor_name = pair.first;
                 if (!starts_with(tensor_name, prefix))
                     continue;
                 if (tensor_name.find("guidance_in.in_layer.weight") != std::string::npos) {
-                    // not schnell
                     flux_params.guidance_embed = true;
                 }
                 if (tensor_name.find("__x0__") != std::string::npos) {
@@ -1344,13 +1340,30 @@ namespace Flux {
                         flux_params.depth_single_blocks = block_depth + 1;
                     }
                 }
+                if (ends_with(tensor_name, "txt_in.weight")) {
+                    flux_params.context_in_dim = pair.second.ne[0];
+                    flux_params.hidden_size    = pair.second.ne[1];
+                }
+                if (ends_with(tensor_name, "single_blocks.0.norm.key_norm.scale")) {
+                    head_dim = pair.second.ne[0];
+                }
+                if (ends_with(tensor_name, "double_blocks.0.txt_attn.norm.key_norm.scale")) {
+                    head_dim = pair.second.ne[0];
+                }
             }
 
-            LOG_INFO("Flux blocks: %d double, %d single", flux_params.depth, flux_params.depth_single_blocks);
+            flux_params.num_heads = static_cast<int>(flux_params.hidden_size / head_dim);
+
+            LOG_INFO("flux: depth = %d, depth_single_blocks = %d, guidance_embed = %s, context_in_dim = %" PRId64
+                     ", hidden_size = %" PRId64 ", num_heads = %d",
+                     flux_params.depth,
+                     flux_params.depth_single_blocks,
+                     flux_params.guidance_embed ? "true" : "false",
+                     flux_params.context_in_dim,
+                     flux_params.hidden_size,
+                     flux_params.num_heads);
             if (flux_params.is_chroma) {
                 LOG_INFO("Using pruned modulation (Chroma)");
-            } else if (!flux_params.guidance_embed) {
-                LOG_INFO("Flux guidance is disabled (Schnell mode)");
             }
 
             flux = Flux(flux_params);

ggml_extend.hpp

@@ -1348,7 +1348,8 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_attention_ext(struct ggml_context
         v = ggml_reshape_3d(ctx, v, L_k, d_head, n_kv_head * N);   // [N * n_kv_head, d_head, L_k]
 
         auto kq = ggml_mul_mat(ctx, k, q);  // [N * n_head, L_q, L_k]
-        kq      = ggml_scale_inplace(ctx, kq, scale);
+        ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
+        kq = ggml_scale_inplace(ctx, kq, scale);
         if (mask) {
             kq = ggml_add_inplace(ctx, kq, mask);
         }

llm.hpp

@@ -837,7 +837,8 @@ namespace LLM {
 
         struct ggml_tensor* forward(GGMLRunnerContext* ctx,
                                     struct ggml_tensor* x,
-                                    struct ggml_tensor* input_pos) {
+                                    struct ggml_tensor* input_pos,
+                                    struct ggml_tensor* attention_mask = nullptr) {
             // x: [N, n_token, hidden_size]
             int64_t n_token = x->ne[1];
             int64_t N       = x->ne[2];
@@ -880,7 +881,7 @@ namespace LLM {
             k = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, k, 0, 2, 1, 3));  // [N, num_kv_heads, n_token, head_dim]
             k = ggml_reshape_3d(ctx->ggml_ctx, k, k->ne[0], k->ne[1], k->ne[2] * k->ne[3]);      // [N*num_kv_heads, n_token, head_dim]
 
-            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, nullptr, true, true, false);  // [N, n_token, hidden_size]
+            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, attention_mask, false, true, false);  // [N, n_token, hidden_size]
 
             x = out_proj->forward(ctx, x);  // [N, n_token, hidden_size]
             return x;
@@ -898,7 +899,8 @@ namespace LLM {
 
         struct ggml_tensor* forward(GGMLRunnerContext* ctx,
                                     struct ggml_tensor* x,
-                                    struct ggml_tensor* input_pos) {
+                                    struct ggml_tensor* input_pos,
+                                    struct ggml_tensor* attention_mask = nullptr) {
             // x: [N, n_token, hidden_size]
             auto self_attn                = std::dynamic_pointer_cast<Attention>(blocks["self_attn"]);
             auto mlp                      = std::dynamic_pointer_cast<MLP>(blocks["mlp"]);
@@ -907,7 +909,7 @@ namespace LLM {
 
             auto residual = x;
             x             = input_layernorm->forward(ctx, x);
-            x             = self_attn->forward(ctx, x, input_pos);
+            x             = self_attn->forward(ctx, x, input_pos, attention_mask);
             x             = ggml_add_inplace(ctx->ggml_ctx, x, residual);
 
             residual = x;
@@ -936,6 +938,7 @@ namespace LLM {
         struct ggml_tensor* forward(GGMLRunnerContext* ctx,
                                     struct ggml_tensor* input_ids,
                                     struct ggml_tensor* input_pos,
+                                    struct ggml_tensor* attention_mask,
                                     std::vector<std::pair<int, ggml_tensor*>> image_embeds,
                                     std::set<int> out_layers) {
             // input_ids: [N, n_token]
@@ -990,7 +993,7 @@ namespace LLM {
             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);
+                x = block->forward(ctx, x, input_pos, attention_mask);
                 if (out_layers.find(i + 1) != out_layers.end()) {
                     intermediate_outputs.push_back(x);
                 }
@@ -1036,12 +1039,13 @@ namespace LLM {
         struct ggml_tensor* forward(GGMLRunnerContext* ctx,
                                     struct ggml_tensor* input_ids,
                                     struct ggml_tensor* input_pos,
+                                    struct ggml_tensor* attention_mask,
                                     std::vector<std::pair<int, ggml_tensor*>> image_embeds,
                                     std::set<int> out_layers) {
             // input_ids: [N, n_token]
             auto model = std::dynamic_pointer_cast<TextModel>(blocks["model"]);
 
-            auto x = model->forward(ctx, input_ids, input_pos, image_embeds, out_layers);
+            auto x = model->forward(ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers);
             return x;
         }
 
@@ -1063,6 +1067,7 @@ namespace LLM {
         LLM model;
 
         std::vector<int> input_pos_vec;
+        std::vector<float> attention_mask_vec;
         std::vector<float> window_mask_vec;
         std::vector<int> window_index_vec;
         std::vector<int> window_inverse_index_vec;
@@ -1157,9 +1162,10 @@ namespace LLM {
         struct ggml_tensor* forward(GGMLRunnerContext* ctx,
                                     struct ggml_tensor* input_ids,
                                     struct ggml_tensor* input_pos,
+                                    struct ggml_tensor* attention_mask,
                                     std::vector<std::pair<int, ggml_tensor*>> image_embeds,
                                     std::set<int> out_layers) {
-            auto hidden_states = model.forward(ctx, input_ids, input_pos, image_embeds, out_layers);  // [N, n_token, hidden_size]
+            auto hidden_states = model.forward(ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers);  // [N, n_token, hidden_size]
             return hidden_states;
         }
 
@@ -1174,6 +1180,7 @@ namespace LLM {
         }
 
         struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
+                                        struct ggml_tensor* attention_mask,
                                         std::vector<std::pair<int, ggml_tensor*>> image_embeds,
                                         std::set<int> out_layers) {
             struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
@@ -1205,9 +1212,26 @@ namespace LLM {
                                                 input_pos_vec.size());
             set_backend_tensor_data(input_pos, input_pos_vec.data());
 
+            if (attention_mask != nullptr) {
+                attention_mask = to_backend(attention_mask);
+            } else {
+                attention_mask_vec.resize(n_tokens * n_tokens);
+                for (int i0 = 0; i0 < n_tokens; i0++) {
+                    for (int i1 = 0; i1 < n_tokens; i1++) {
+                        float value = 0.f;
+                        if (i0 > i1) {
+                            value = -INFINITY;
+                        }
+                        attention_mask_vec[i1 * n_tokens + i0] = value;
+                    }
+                }
+                attention_mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, n_tokens, n_tokens);
+                set_backend_tensor_data(attention_mask, attention_mask_vec.data());
+            }
+
             auto runner_ctx = get_context();
 
-            struct ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, input_pos, image_embeds, out_layers);
+            struct ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers);
 
             ggml_build_forward_expand(gf, hidden_states);
 
@@ -1216,12 +1240,13 @@ namespace LLM {
 
         bool compute(const int n_threads,
                      struct ggml_tensor* input_ids,
+                     struct ggml_tensor* attention_mask,
                      std::vector<std::pair<int, ggml_tensor*>> image_embeds,
                      std::set<int> out_layers,
                      ggml_tensor** output,
                      ggml_context* output_ctx = nullptr) {
             auto get_graph = [&]() -> struct ggml_cgraph* {
-                return build_graph(input_ids, image_embeds, out_layers);
+                return build_graph(input_ids, attention_mask, image_embeds, out_layers);
             };
             return GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
         }
@@ -1525,7 +1550,7 @@ namespace LLM {
                 struct ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
-                model.compute(8, input_ids, image_embeds, {}, &out, work_ctx);
+                model.compute(8, input_ids, nullptr, image_embeds, {}, &out, work_ctx);
                 int64_t t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);
@@ -1565,7 +1590,7 @@ namespace LLM {
                 struct ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
-                model.compute(8, input_ids, {}, {10, 20, 30}, &out, work_ctx);
+                model.compute(8, input_ids, nullptr, {}, {10, 20, 30}, &out, work_ctx);
                 int64_t t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);
@@ -1588,7 +1613,7 @@ namespace LLM {
                 struct ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
-                model.compute(8, input_ids, {}, {35}, &out, work_ctx);
+                model.compute(8, input_ids, nullptr, {}, {35}, &out, work_ctx);
                 int64_t t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);
@@ -1611,7 +1636,7 @@ namespace LLM {
                 struct ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
-                model.compute(8, input_ids, {}, {}, &out, work_ctx);
+                model.compute(8, input_ids, nullptr, {}, {}, &out, work_ctx);
                 int64_t t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);

model.cpp

@@ -1034,10 +1034,13 @@ SDVersion ModelLoader::get_sd_version() {
 
     bool is_xl                       = false;
     bool is_flux                     = false;
+    bool is_flux2                    = false;
+    bool has_single_block_47         = false;
     bool is_wan                      = false;
     int64_t patch_embedding_channels = 0;
     bool has_img_emb                 = false;
     bool has_middle_block_1          = false;
+    bool has_output_block_311        = false;
     bool has_output_block_71         = false;
 
     for (auto& [name, tensor_storage] : tensor_storage_map) {
@@ -1055,7 +1058,10 @@ SDVersion ModelLoader::get_sd_version() {
                 return VERSION_QWEN_IMAGE;
             }
             if (tensor_storage.name.find("model.diffusion_model.double_stream_modulation_img.lin.weight") != std::string::npos) {
-                return VERSION_FLUX2;
+                is_flux2 = true;
+            }
+            if (tensor_storage.name.find("single_blocks.47.linear1.weight") != std::string::npos) {
+                has_single_block_47 = true;
             }
             if (tensor_storage.name.find("model.diffusion_model.double_blocks.0.img_mlp.gate_proj.weight") != std::string::npos) {
                 return VERSION_OVIS_IMAGE;
@@ -1095,6 +1101,9 @@ SDVersion ModelLoader::get_sd_version() {
             tensor_storage.name.find("unet.mid_block.resnets.1.") != std::string::npos) {
             has_middle_block_1 = true;
         }
+        if (tensor_storage.name.find("model.diffusion_model.output_blocks.3.1.transformer_blocks.1") != std::string::npos) {
+            has_output_block_311 = true;
+        }
         if (tensor_storage.name.find("model.diffusion_model.output_blocks.7.1") != std::string::npos) {
             has_output_block_71 = true;
         }
@@ -1133,12 +1142,15 @@ SDVersion ModelLoader::get_sd_version() {
             return VERSION_SDXL_PIX2PIX;
         }
         if (!has_middle_block_1) {
+            if (!has_output_block_311) {
+                return VERSION_SDXL_VEGA;
+            }
             return VERSION_SDXL_SSD1B;
         }
         return VERSION_SDXL;
     }
 
-    if (is_flux) {
+    if (is_flux && !is_flux2) {
         if (input_block_weight.ne[0] == 384) {
             return VERSION_FLUX_FILL;
         }
@@ -1151,6 +1163,13 @@ SDVersion ModelLoader::get_sd_version() {
         return VERSION_FLUX;
     }
 
+    if (is_flux2) {
+        if (has_single_block_47) {
+            return VERSION_FLUX2;
+        }
+        return VERSION_FLUX2_KLEIN;
+    }
+
     if (token_embedding_weight.ne[0] == 768) {
         if (is_inpaint) {
             return VERSION_SD1_INPAINT;

model.h

@@ -32,6 +32,7 @@ enum SDVersion {
     VERSION_SDXL,
     VERSION_SDXL_INPAINT,
     VERSION_SDXL_PIX2PIX,
+    VERSION_SDXL_VEGA,
     VERSION_SDXL_SSD1B,
     VERSION_SVD,
     VERSION_SD3,
@@ -45,6 +46,7 @@ enum SDVersion {
     VERSION_WAN2_2_TI2V,
     VERSION_QWEN_IMAGE,
     VERSION_FLUX2,
+    VERSION_FLUX2_KLEIN,
     VERSION_Z_IMAGE,
     VERSION_OVIS_IMAGE,
     VERSION_COUNT,
@@ -65,7 +67,7 @@ static inline bool sd_version_is_sd2(SDVersion version) {
 }
 
 static inline bool sd_version_is_sdxl(SDVersion version) {
-    if (version == VERSION_SDXL || version == VERSION_SDXL_INPAINT || version == VERSION_SDXL_PIX2PIX || version == VERSION_SDXL_SSD1B) {
+    if (version == VERSION_SDXL || version == VERSION_SDXL_INPAINT || version == VERSION_SDXL_PIX2PIX || version == VERSION_SDXL_SSD1B || version == VERSION_SDXL_VEGA) {
         return true;
     }
     return false;
@@ -100,7 +102,7 @@ static inline bool sd_version_is_flux(SDVersion version) {
 }
 
 static inline bool sd_version_is_flux2(SDVersion version) {
-    if (version == VERSION_FLUX2) {
+    if (version == VERSION_FLUX2 || version == VERSION_FLUX2_KLEIN) {
         return true;
     }
     return false;

stable-diffusion.cpp

@@ -35,6 +35,7 @@ const char* model_version_to_str[] = {
     "SDXL",
     "SDXL Inpaint",
     "SDXL Instruct-Pix2Pix",
+    "SDXL (Vega)",
     "SDXL (SSD1B)",
     "SVD",
     "SD3.x",
@@ -48,6 +49,7 @@ const char* model_version_to_str[] = {
     "Wan 2.2 TI2V",
     "Qwen Image",
     "Flux.2",
+    "Flux.2 klein",
     "Z-Image",
     "Ovis Image",
 };
@@ -622,7 +624,7 @@ public:
                         LOG_INFO("Using Conv2d direct in the vae model");
                         first_stage_model->set_conv2d_direct_enabled(true);
                     }
-                    if (version == VERSION_SDXL &&
+                    if (sd_version_is_sdxl(version) &&
                         (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 || sd_ctx_params->force_sdxl_vae_conv_scale)) {
                         float vae_conv_2d_scale = 1.f / 32.f;
                         LOG_WARN(

unet.hpp

@@ -201,6 +201,9 @@ public:
             num_head_channels     = 64;
             num_heads             = -1;
             use_linear_projection = true;
+            if (version == VERSION_SDXL_VEGA) {
+                transformer_depth = {1, 1, 2};
+            }
         } else if (version == VERSION_SVD) {
             in_channels           = 8;
             out_channels          = 4;
@@ -319,7 +322,7 @@ public:
         }
         if (!tiny_unet) {
             blocks["middle_block.0"] = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, ch));
-            if (version != VERSION_SDXL_SSD1B) {
+            if (version != VERSION_SDXL_SSD1B && version != VERSION_SDXL_VEGA) {
                 blocks["middle_block.1"] = std::shared_ptr<GGMLBlock>(get_attention_layer(ch,
                                                                                           n_head,
                                                                                           d_head,
@@ -520,7 +523,7 @@ public:
         // middle_block
         if (!tiny_unet) {
             h = resblock_forward("middle_block.0", ctx, h, emb, num_video_frames);  // [N, 4*model_channels, h/8, w/8]
-            if (version != VERSION_SDXL_SSD1B) {
+            if (version != VERSION_SDXL_SSD1B && version != VERSION_SDXL_VEGA) {
                 h = attention_layer_forward("middle_block.1", ctx, h, context, num_video_frames);  // [N, 4*model_channels, h/8, w/8]
                 h = resblock_forward("middle_block.2", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
             }