feat(gallery): add MOSS-Image and wire tokenizer option (#12757)

stable-diffusion.cpp PR #2148 added MOSS-Image support, which needs an
external tokenizer file passed via the --tokenizer CLI arg. The C++ binding
(gosd.cpp) never parsed a tokenizer option nor set ctx_params.tokenizer, so
MOSS-Image could not load through LocalAI.

- Wire the "tokenizer" option in gosd.cpp (parse + assign to ctx_params)
- Resolve "tokenizer:..." paths against the model directory in gosd.go,
  the same way *_path options are resolved
- Bump the vendored stable-diffusion.cpp to 68285df (moss-image merge)
- Add moss-image-base and moss-image-sft gallery entries (int8_convrot)

FLUX.3 Image Klein is already in the gallery (added in #12736).

Co-authored-by: Ettore Di Giacinto <mudler@localai.io>

backend/go/stablediffusion-ggml/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # stablediffusion.cpp (ggml)
 STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=8cd731e7b4667eb8f40a153315b9a7125d09a58d
+STABLEDIFFUSION_GGML_VERSION?=68285df319457e2d79310b48513e4ba1bf8ebe9a
 
 CMAKE_ARGS+=-DGGML_MAX_NAME=128
 

backend/go/stablediffusion-ggml/cpp/gosd.cpp

@@ -395,6 +395,7 @@ int load_model(const char *model, char *model_path, char* options[], int threads
     const char *photo_maker_path = "";
     const char *pulid_weights_path = "";
     const char *tensor_type_rules = "";
+    const char *tokenizer_path = "";
     char *lora_dir = model_path;
 
     // Upstream backend/parameter placement specs (see docs/.../stablediffusion).
@@ -503,6 +504,7 @@ int load_model(const char *model, char *model_path, char* options[], int threads
         if (!strcmp(optname, "photo_maker_path")) photo_maker_path = strdup(optval);
         if (!strcmp(optname, "pulid_weights_path")) pulid_weights_path = strdup(optval);
         if (!strcmp(optname, "tensor_type_rules")) tensor_type_rules = strdup(optval);
+        if (!strcmp(optname, "tokenizer")) tokenizer_path = strdup(optval);
 
         // Backend / parameter placement specs (see prepare_backend_assignments
         // in the upstream CLI). These compose with the legacy keep_*_on_cpu /
@@ -619,6 +621,7 @@ int load_model(const char *model, char *model_path, char* options[], int threads
     ctx_params.photo_maker_path = photo_maker_path;
     if (strlen(pulid_weights_path) > 0) ctx_params.pulid_weights_path = pulid_weights_path;
     ctx_params.tensor_type_rules = tensor_type_rules;
+    if (strlen(tokenizer_path) > 0) ctx_params.tokenizer = tokenizer_path;
     ctx_params.n_threads = n_threads;
     ctx_params.rng_type = rng_type;
     if (wtype != SD_TYPE_COUNT) ctx_params.wtype = wtype;

backend/go/stablediffusion-ggml/gosd.go

@@ -359,7 +359,7 @@ func (sd *SDGGML) Load(opts *pb.ModelOptions) error {
 		}
 
 		// If it's an option path, we resolve absolute path from the model path
-		if strings.Contains(op, ":") && strings.Contains(op, "path") {
+		if strings.Contains(op, ":") && (strings.Contains(op, "path") || strings.Contains(op, "tokenizer")) {
 			data := strings.Split(op, ":")
 			if err := utils.VerifyPath(data[1], opts.ModelPath); err == nil {
 				data[1] = filepath.Join(opts.ModelPath, data[1])

gallery/index.yaml

@@ -52091,6 +52091,97 @@
     - filename: stablediffusion-cpp/models/Qwen3VL-4B-Instruct-Q4_K_M.gguf
       sha256: 66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a
       uri: https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf
+- name: moss-image-base
+  url: github:mudler/LocalAI/gallery/flux-ggml.yaml@master
+  urls:
+    - https://huggingface.co/Comfy-Org/MOSS-Image
+    - https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base
+  description: |
+    MOSS-Image-Base is a 5B diffusion transformer image generation model from OpenMOSS-Team. It uses Qwen3.5 9B as the LLM text encoder, the FLUX.2 VAE, and an external tokenizer. The Base variant supports text-to-image generation with classifier-free guidance (default: CFG 3.0, 1024x1024).
+
+    This entry uses the int8_convrot quantization of the diffusion model (~5.4 GB).
+  license: apache-2.0
+  icon: https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base/resolve/main/assets/teaser.webp
+  tags:
+    - moss-image
+    - text-to-image
+    - image-generation
+    - diffusion
+    - 5b
+    - quantized
+    - safetensors
+  last_checked: "2026-10-18"
+  overrides:
+    cfg_scale: 3
+    options:
+      - diffusion_model
+      - vae_path:stablediffusion-cpp/models/flux2-vae.safetensors
+      - sampler:euler
+      - llm_path:stablediffusion-cpp/models/Qwen3.5-9B-Q4_K_M.gguf
+      - tokenizer:stablediffusion-cpp/models/moss-image-tokenizer.json
+      - offload_params_to_cpu:true
+      - vae_decode_only:false
+    parameters:
+      model: stablediffusion-cpp/models/moss_image_base_int8_convrot.safetensors
+    step: 20
+  files:
+    - filename: stablediffusion-cpp/models/moss_image_base_int8_convrot.safetensors
+      uri: https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/diffusion_models/moss_image_base_int8_convrot.safetensors
+    - filename: stablediffusion-cpp/models/flux2-vae.safetensors
+      sha256: d64f3a68e1cc4f9f4e29b6e0da38a0204fe9a49f2d4053f0ec1fa1ca02f9c4b5
+      uri: https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/vae/flux2-vae.safetensors
+    - filename: stablediffusion-cpp/models/Qwen3.5-9B-Q4_K_M.gguf
+      sha256: 03b74727a860a56338e042c4420bb3f04b2fec5734175f4cb9fa853daf52b7e8
+      uri: https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/resolve/main/Qwen3.5-9B-Q4_K_M.gguf
+    - filename: stablediffusion-cpp/models/moss-image-tokenizer.json
+      sha256: 06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+      uri: https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base/resolve/main/tokenizer/tokenizer.json
+- name: moss-image-sft
+  url: github:mudler/LocalAI/gallery/flux-ggml.yaml@master
+  urls:
+    - https://huggingface.co/Comfy-Org/MOSS-Image
+    - https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT
+  description: |
+    MOSS-Image-SFT is the supervised fine-tuned variant of MOSS-Image, a 5B diffusion transformer image generation model from OpenMOSS-Team. It uses Qwen3.5 9B as the LLM text encoder, the FLUX.2 VAE, and an external tokenizer. The SFT variant produces higher-quality outputs with classifier-free guidance (default: CFG 3.0, 1024x1024).
+
+    This entry uses the int8_convrot quantization of the diffusion model (~5.4 GB).
+  license: apache-2.0
+  icon: https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base/resolve/main/assets/teaser.webp
+  tags:
+    - moss-image
+    - moss-image-sft
+    - text-to-image
+    - image-generation
+    - diffusion
+    - 5b
+    - quantized
+    - safetensors
+  last_checked: "2026-10-18"
+  overrides:
+    cfg_scale: 3
+    options:
+      - diffusion_model
+      - vae_path:stablediffusion-cpp/models/flux2-vae.safetensors
+      - sampler:euler
+      - llm_path:stablediffusion-cpp/models/Qwen3.5-9B-Q4_K_M.gguf
+      - tokenizer:stablediffusion-cpp/models/moss-image-tokenizer.json
+      - offload_params_to_cpu:true
+      - vae_decode_only:false
+    parameters:
+      model: stablediffusion-cpp/models/moss_image_sft_int8_convrot.safetensors
+    step: 20
+  files:
+    - filename: stablediffusion-cpp/models/moss_image_sft_int8_convrot.safetensors
+      uri: https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/diffusion_models/moss_image_sft_int8_convrot.safetensors
+    - filename: stablediffusion-cpp/models/flux2-vae.safetensors
+      sha256: d64f3a68e1cc4f9f4e29b6e0da38a0204fe9a49f2d4053f0ec1fa1ca02f9c4b5
+      uri: https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/vae/flux2-vae.safetensors
+    - filename: stablediffusion-cpp/models/Qwen3.5-9B-Q4_K_M.gguf
+      sha256: 03b74727a860a56338e042c4420bb3f04b2fec5734175f4cb9fa853daf52b7e8
+      uri: https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/resolve/main/Qwen3.5-9B-Q4_K_M.gguf
+    - filename: stablediffusion-cpp/models/moss-image-tokenizer.json
+      sha256: 06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+      uri: https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base/resolve/main/tokenizer/tokenizer.json
 - name: Z-Image-Turbo
   url: github:mudler/LocalAI/gallery/z-image-ggml.yaml@master
   urls: