diff --git a/CHANGELOG.md b/CHANGELOG.md
index aacd64ce1..29ab781b9 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,9 +4,10 @@
 
 ### Highlights for 2026-07-02
 
-Service-pack update with several fixes and quality-of-life improvements  
-Plus few new models: **Krea 2**, **Photoroom PRXPixel**, **FLUX.2 Klein 9B KV**  
+Service-pack update with number of fixes and quality-of-life improvements  
+Plus few new models: **Krea 2**, **Photoroom PRXPixel**, **FLUX.2 Klein 9B KV** and some new community models  
 And **SDNQ** improvements: now with *NPU* support and its own native *attention* kernels!  
+Also couple of *experimental* features: see below for details...  
 
 [Home](https://vladmandic.github.io/sdnext/) | [ChangeLog](https://github.com/vladmandic/automatic/blob/master/CHANGELOG.md) | [Docs](https://vladmandic.github.io/sdnext-docs/) | [Discord](https://discord.com/invite/sd-next-federal-batch-inspectors-1101998836328697867) | [Sponsor](https://github.com/sponsors/vladmandic)  
 
@@ -34,6 +35,7 @@ And **SDNQ** improvements: now with *NPU* support and its own native *attention*
   - add option: *model loading -> attempt to load incomplete model*  
     disabled by default, attempts to load model by mapping it to known model even if some components are missing  
     for example: if you place bare unet/dit finetune into stable-diffusion folder  
+  - prompt encode caching for pass-through text-encoders
 - **UI**
   - dynamic visibility of image controls
   - improve main panel positioning: *portrait/landscape*
@@ -44,6 +46,9 @@ And **SDNQ** improvements: now with *NPU* support and its own native *attention*
 - **Internal**
   - delay init of video models
 - **Experimental**
+  - support for **openai interface** for llm  
+    in *prompt enhance* enable openai interface and when llm is loaded,  
+    sdnext will start a local openai-compatible server on usual endpoints (e.g. `/v1/completions`, `/v1/chat/completions`, etc.)  
   - support for [pruna](https://docs.pruna.ai/en/stable/compression.html) swiss-army-knife of model compression, caching and optimization  
     see *settings -> model compile* for options  
     *note*: pruna options compatibility varies greatly depending on platform, gpu, torch and model used  
@@ -65,6 +70,7 @@ And **SDNQ** improvements: now with *NPU* support and its own native *attention*
   - measure: handle current kanvas stage
   - model metadata: handle invalid metadata and strip workflows
   - mps: install `torchsde` as requirement
+  - nunchaku: gate for `cuda` only
   - onnxruntime: handle invalid version
   - onnxruntime: mark all import paths as non-critical
   - options: handle compatibility options
diff --git a/TODO.md b/TODO.md
index efb27fe06..00a6b31bc 100644
--- a/TODO.md
+++ b/TODO.md
@@ -31,6 +31,7 @@
 - Integrate natural language image search
   - [ImageDB](https://github.com/vladmandic/imagedb)
 - Unify *huggingface* and *diffusers* model folders
+- [QuantFunc](https://huggingface.co/QuantFunc/Klein-9B-Series)
 
 ### OnHold
 
diff --git a/data/reference-community.json b/data/reference-community.json
index bbc1de240..a2ac6b8a2 100644
--- a/data/reference-community.json
+++ b/data/reference-community.json
@@ -169,5 +169,19 @@
     "desc": "Anima 1.0 Base pre-merged with several LoRAs and quantized to uint4 using SDNQ with Hadamard. Flexible as it can be used with and without guidance.",
     "date": "2026 July",
     "size": 2.22
+  },
+  "FLUX.2 Klein 9B KV Merge sdnq-hadamard-svd": {
+    "path": "vladmandic/Flux.2-Klein-9B-KV-Merge-sdnq-hadamard-uint4",
+    "preview": "black-forest-labs--FLUX.2-klein-9b-kv.jpg",
+    "desc": "FLUX.2 Klein 9B KV pre-merged with several LoRAs and quantized to uint4 using SDNQ with Hadamard.",
+    "size": 11.67,
+    "date": "2026 July"
+  },
+  "Krea 2 Turbo Merge sdnq-hadamard-uint4": {
+    "path": "vladmandic/Krea-2-Turbo-Merge-sdnq-hadamard-uint4",
+    "preview": "CalamitousFelicitousness--Krea-2-Turbo-Diffusers.jpg",
+    "desc": "Krea 2 Turbo pre-merged with several LoRAs and quantized to uint4 using SDNQ with Hadamard.",
+    "size": 10.55,
+    "date": "2026 July"
   }
 }
diff --git a/modules/api/models.py b/modules/api/models.py
index 5ed7743fd..cae3a44d5 100644
--- a/modules/api/models.py
+++ b/modules/api/models.py
@@ -375,6 +375,7 @@ class ReqPromptEnhance(BaseModel):
     process_words: Optional[str] = Field(title="Banned words", default=None, description="List of words to process")
     semantic_threshold: Optional[float] = Field(title="Semantic threshold", default=None, description="Semantic similarity threshold for processed words")
     embedding_similarity: Optional[float] = Field(title="Embedding similarity", default=None, description="Embedding similarity threshold for processed words")
+    use_openai: Optional[bool] = Field(title="Use OpenAI", default=False, description="Use OpenAI API for model access")
 
 class ResPromptEnhance(BaseModel):
     prompt: str = Field(title="Prompt", description="Enhanced prompt")
diff --git a/modules/api/process.py b/modules/api/process.py
index 1abf3950b..ecb1659c7 100644
--- a/modules/api/process.py
+++ b/modules/api/process.py
@@ -247,6 +247,7 @@ class APIProcess:
                 process_words=req.process_words,
                 semantic_threshold=req.semantic_threshold,
                 embedding_similarity=req.embedding_similarity,
+                use_openai=req.use_openai
             )
         elif req.type == 'video':
             from modules.ui_video_vlm import enhance_prompt
diff --git a/modules/openai/serve.py b/modules/openai/serve.py
index 5c3fa3b2f..f46f015b8 100644
--- a/modules/openai/serve.py
+++ b/modules/openai/serve.py
@@ -17,6 +17,7 @@ class OpenAIServer:
         processor=None,
         host: str = "127.0.0.1",
         port: int = 8888,
+        server: Optional[uvicorn.Server] = None,
         max_context_tokens: Optional[int] = None,
         max_new_tokens: Optional[int] = None,
         stream: Optional[bool] = None,
@@ -24,7 +25,7 @@ class OpenAIServer:
         top_p: Optional[float] = None,
         top_k: Optional[int] = None,
         repetition_penalty: Optional[float] = None,
-        api_key: Optional[str] = None
+        api_key: Optional[str] = None,
     ):
         self.model = model
         self.tokenizer = tokenizer
@@ -50,46 +51,55 @@ class OpenAIServer:
         log.info(f"OpenAI: {self.model_info}")
         attention = get_attention_config(self)
         log.debug(f'OpenAI: {attention}')
-        self.api_key = api_key
-        self._security = HTTPBearer(auto_error=False)
-        self._is_running = False
-        self._lock = threading.Lock()
+
+        if server:
+            self.use_server = True
+            self.app = server
+            self.server = server
+            self._is_running = True
+        else:
+            self.use_server = False
+            self.app = FastAPI(title="SD.Next OpenAI-compatible LLM Server", version="1.0")
+            self.server: Optional[uvicorn.Server] = None
+            self._is_running = False
         self._startup_event = threading.Event()
-        self.server: Optional[uvicorn.Server] = None
+        self.api_key = api_key
+        self._lock = threading.Lock()
+        self._security = HTTPBearer(auto_error=False)
         self.thread: Optional[threading.Thread] = None
-        self.app = FastAPI(title="SD.Next OpenAI-compatible LLM Server", version="1.0")
         setup_routes(self)
 
     def start(self, timeout_seconds: float = 10.0):
         """Spawns the serving interface safely using an active background thread worker."""
+        if self._is_running:
+            # log.warning("OpenAI: Server('already running')")
+            return
         with self._lock:
-            if self._is_running:
-                log.warning("OpenAI: Server('already running')")
-                return
-            self._startup_event.clear()
-            config = uvicorn.Config(
-                app=self.app, host=self.host, port=self.port, log_level="info", loop="asyncio", workers=1
-            )
-            self.server = uvicorn.Server(config)
-            self.server.install_signal_handlers = lambda *args, **kwargs: None
-            original_startup = self.server.startup
+            if self.server is None:
+                self._startup_event.clear()
+                config = uvicorn.Config(app=self.app, host=self.host, port=self.port, log_level="info", loop="asyncio", workers=1)
+                self.server = uvicorn.Server(config)
+                self.server.install_signal_handlers = lambda *args, **kwargs: None
+                original_startup = self.server.startup
 
-            async def patched_startup(*args, **kwargs):
-                await original_startup(*args, **kwargs)
-                self._startup_event.set()
+                async def patched_startup(*args, **kwargs):
+                    await original_startup(*args, **kwargs)
+                    self._startup_event.set()
 
-            self.server.startup = patched_startup
-            self.thread = threading.Thread(target=self.server.run, name="TransformersServeWorkerThread", daemon=True)
-            self.thread.start()
-        if not self._startup_event.wait(timeout=timeout_seconds):
-            self.stop()
-            raise TimeoutError("OpenAI: init timeout")
+                self.server.startup = patched_startup
+                self.thread = threading.Thread(target=self.server.run, name="TransformersServeWorkerThread", daemon=True)
+                self.thread.start()
+            if not self._startup_event.wait(timeout=timeout_seconds):
+                self.stop()
+                raise TimeoutError("OpenAI: init timeout")
         self._is_running = True
         url = f"http://{self.host}:{self.port}/v1"
         log.info(f"OpenAI: Server(url={url})")
 
     def stop(self, timeout_seconds: float = 5.0):
         """Safely winds down network sockets and detached worker threads."""
+        if self.use_server:
+            return
         with self._lock:
             if not self._is_running or not self.server:
                 return
@@ -98,5 +108,7 @@ class OpenAIServer:
                 self.thread.join(timeout=timeout_seconds)
             self.server = None
             self.thread = None
+            self.server = None
+            self.thread = None
             self._is_running = False
             log.info("OpenAI: Server(None)")
diff --git a/modules/sd_hijack_te.py b/modules/sd_hijack_te.py
index a9f7f9d29..30d86c087 100644
--- a/modules/sd_hijack_te.py
+++ b/modules/sd_hijack_te.py
@@ -4,6 +4,40 @@ from modules import shared, errors, timer, sd_models
 from modules.logger import log
 
 
+class PromptCache:
+    def __init__(self):
+        self.cache = {}
+        self.id = None
+        self.max = 16
+
+    def get(self, prompt):
+        if self.id != id(shared.sd_model):
+            self.cache.clear()
+            self.id = id(shared.sd_model)
+            log.debug(f'Encode: prompt cache activate id={self.id} depth={len(self.cache)}')
+        if (isinstance(prompt, list) and len(prompt) == 1 and isinstance(prompt[0], str)):
+            cached = self.cache.get(prompt[0], None)
+        elif isinstance(prompt, str):
+            cached = self.cache.get(prompt, None)
+        else:
+            cached = None
+        if cached:
+            log.debug(f'Encode: prompt="{prompt}" cache={len(self.cache)} hit')
+        return cached
+
+    def set(self, prompt, encoded):
+        if len(self.cache) >= self.max:
+            oldest_key = next(iter(self.cache))
+            del self.cache[oldest_key]
+        if (isinstance(prompt, list) and len(prompt) == 1 and isinstance(prompt[0], str)):
+            self.cache[prompt[0]] = encoded
+        elif isinstance(prompt, str):
+            self.cache[prompt] = encoded
+
+
+prompt_cache = PromptCache()
+
+
 def hijack_encode_prompt(*args, **kwargs):
     jobid = shared.state.begin('TE Encode')
     t0 = time.time()
@@ -17,8 +51,8 @@ def hijack_encode_prompt(*args, **kwargs):
         if prompt is None and len(args_copy) > 0:
             prompt = args_copy[0]
             patch_prompt = True
+        prompt = [p.strip(", \n") if isinstance(p, str) else p for p in prompt] if isinstance(prompt, list) else prompt
         res = prompt
-        log.debug(f'Encode: prompt="{prompt}" hijack=True')
 
         if hasattr(shared.sd_model, 'before_prompt_encode'):
             log.debug(f'Encode: prompt="{prompt}" op=before')
@@ -26,8 +60,16 @@ def hijack_encode_prompt(*args, **kwargs):
             if patch_prompt:
                 args_copy[0] = res
 
-        if hasattr(shared.sd_model, 'orig_encode_prompt'):
-            res = shared.sd_model.orig_encode_prompt(*args_copy, **kwargs)
+        cached = prompt_cache.get(prompt)
+        if cached is not None:
+            res = cached
+        else:
+            log.debug(f'Encode: prompt="{prompt}" hijack=True')
+            if hasattr(shared.sd_model, 'orig_encode_prompt'):
+                res = shared.sd_model.orig_encode_prompt(*args_copy, **kwargs)
+            else:
+                res = shared.sd_model.encode_prompt(*args_copy, **kwargs)
+            prompt_cache.set(prompt, res)
 
         if hasattr(shared.sd_model, 'after_prompt_encode'):
             log.debug(f'Encode: prompt="{prompt}" op=after')
diff --git a/modules/ui_extra_networks.py b/modules/ui_extra_networks.py
index 690d004c0..d65e549fd 100644
--- a/modules/ui_extra_networks.py
+++ b/modules/ui_extra_networks.py
@@ -15,7 +15,7 @@ from collections import OrderedDict
 import gradio as gr
 from PIL import Image
 from starlette.responses import FileResponse, JSONResponse
-from modules import paths, shared, files_cache, errors, infotext, ui_symbols, ui_components, modelstats
+from modules import paths, shared, devices, files_cache, errors, infotext, ui_symbols, ui_components, modelstats
 from modules.logger import log
 from modules.json_helpers import writefile
 
@@ -321,7 +321,8 @@ class ExtraNetworksPage:
             subdirs['Base'] = 1
             subdirs['Distilled'] = 1
             subdirs['Quantized'] = 1
-            subdirs['Nunchaku'] = 1
+            if devices.backend == 'cuda':
+                subdirs['Nunchaku'] = 1
             subdirs['Community'] = 1
             subdirs['Cloud'] = 1
             subdirs[diffusers_base] = 1
@@ -355,7 +356,9 @@ class ExtraNetworksPage:
                 continue
             if subdir in ['All', 'Local', 'Diffusers']:
                 style = 'network-local'
-            elif subdir in ['Base', 'Reference', 'Distilled', 'Quantized', 'Nunchaku', 'Community', 'Cloud']:
+            elif subdir in ['Base', 'Reference', 'Distilled', 'Quantized', 'Community', 'Cloud']:
+                style = 'network-reference'
+            elif subdir in ['Nunchaku'] and devices.backend == 'cuda':
                 style = 'network-reference'
             else:
                 style = 'network-folder'
diff --git a/pipelines/model_flux2_klein.py b/pipelines/model_flux2_klein.py
index b1599f18d..a6eb649bd 100644
--- a/pipelines/model_flux2_klein.py
+++ b/pipelines/model_flux2_klein.py
@@ -20,7 +20,7 @@ def load_flux2_klein(checkpoint_info, diffusers_load_config=None):
     if repo_id is None or repo_id.lower() == 'none':
         return None
 
-    if '-kv' in repo_id:
+    if '-kv' in repo_id.lower():
         cls = diffusers.Flux2KleinKVPipeline
     else:
         cls = diffusers.Flux2KleinPipeline
diff --git a/scripts/prompt_enhance.py b/scripts/prompt_enhance.py
index 63cfe3dd0..c00357b71 100644
--- a/scripts/prompt_enhance.py
+++ b/scripts/prompt_enhance.py
@@ -302,6 +302,7 @@ class PromptEnhanceScript(scripts_manager.Script):
     processor: transformers.AutoProcessor = None
     tokenizer: transformers.AutoTokenizer = None
     busy: bool = False
+    server = None
     options = Options()
 
     def title(self):
@@ -316,7 +317,7 @@ class PromptEnhanceScript(scripts_manager.Script):
         from modules.sd_models_compile import compile_torch
         self.llm = compile_torch(self.llm, apply_to_components=False, op="LLM")
 
-    def load(self, name:str | None=None, model_repo:str | None=None, model_gguf:str | None=None, model_type:str | None=None, model_file:str | None=None):
+    def load(self, name:str | None=None, use_openai:bool=False, model_repo:str | None=None, model_gguf:str | None=None, model_type:str | None=None, model_file:str | None=None):
         # Strip symbols from display name if present
         name = get_model_repo_from_display(name) if name else self.options.default
         if self.busy:
@@ -419,7 +420,9 @@ class PromptEnhanceScript(scripts_manager.Script):
             log.error(f'Prompt enhance: load {e}')
             if debug_enabled:
                 errors.display(e, 'Prompt enhance')
+
         devices.torch_gc()
+        self.set_openai(enable=use_openai)
         self.busy = False
         return model_repo
 
@@ -430,6 +433,7 @@ class PromptEnhanceScript(scripts_manager.Script):
     def unload(self):
         if self.llm is not None:
             model_name = self.model
+            self.set_openai(enable=False)
             log.debug(f'Prompt enhance: unloading model="{model_name}"')
             deregister_aux('prompt_enhance')
             sd_models.move_model(self.llm, devices.cpu, force=True)
@@ -442,6 +446,21 @@ class PromptEnhanceScript(scripts_manager.Script):
         else:
             log.debug('Prompt enhance: no model loaded')
 
+    def set_openai(self, enable: bool):
+        from modules.openai.serve import OpenAIServer
+        if enable and self.llm is not None and self.tokenizer is not None:
+            self.server = OpenAIServer(
+                model=self.llm,
+                tokenizer=self.tokenizer,
+                host="127.0.0.1",
+                port=8000,
+                server=shared.api.app,
+            )
+            self.server.start()
+        elif self.server is not None:
+            self.server.stop()
+            self.server = None
+
     def clean(self, response, keep_thinking=False, prefill_text='', keep_prefill=False):
         # Handle thinking tags FIRST (before generic tag removal)
         if '<think>' in response or '</think>' in response:
@@ -573,6 +592,7 @@ class PromptEnhanceScript(scripts_manager.Script):
                 process_words:str='',
                 semantic_threshold:float=0.0,
                 embedding_similarity:float=0.0,
+                use_openai:bool=False,
                ):
         # Strip symbols from model name if present
         model = get_model_repo_from_display(model) if model else self.options.default
@@ -599,7 +619,7 @@ class PromptEnhanceScript(scripts_manager.Script):
             time.sleep(0.1)
 
         if not is_cloud_model(model):
-            self.load(model)
+            self.load(model, use_openai=use_openai)
 
         if seed is None or seed == -1:
             random.seed()
@@ -859,7 +879,7 @@ class PromptEnhanceScript(scripts_manager.Script):
             return prompt # Return original full prompt on censorship
         return response
 
-    def apply(self, prompt, image, apply_prompt, llm_model, prompt_system, prompt_prefix, prompt_suffix, min_tokens, max_tokens, do_sample, temperature, repetition_penalty, top_k, top_p, thinking_mode, nsfw_mode, use_vision, prefill_text, keep_prefill, keep_thinking, custom_args, process_words, semantic_threshold, embedding_similarity):
+    def apply(self, prompt, image, apply_prompt, llm_model, prompt_system, prompt_prefix, prompt_suffix, min_tokens, max_tokens, do_sample, temperature, repetition_penalty, top_k, top_p, thinking_mode, nsfw_mode, use_vision, prefill_text, keep_prefill, keep_thinking, custom_args, process_words, semantic_threshold, embedding_similarity, use_openai):
         response = self.enhance(
             prompt=prompt,
             image=image,
@@ -884,6 +904,7 @@ class PromptEnhanceScript(scripts_manager.Script):
             process_words=process_words,
             semantic_threshold=semantic_threshold,
             embedding_similarity=embedding_similarity,
+            use_openai=use_openai,
         )
         if apply_prompt:
             return [response, response]
@@ -918,13 +939,14 @@ class PromptEnhanceScript(scripts_manager.Script):
                 # Set initial state based on whether default model supports vision
                 default_is_vl = is_vision_model(Options.default)
                 use_vision = gr.Checkbox(label='Use vision', value=False, interactive=default_is_vl, elem_id='prompt_enhance_use_vision')
+                use_openai = gr.Checkbox(label='OpenAI interface', value=False, elem_id='prompt_enhance_openai')
             gr.HTML('<br>')
             with gr.Group():
                 with gr.Row():
                     llm_model = gr.Dropdown(label='LLM model', choices=Options.get_model_choices(), value=Options.get_default_display(), interactive=True, allow_custom_value=True, elem_id='prompt_enhance_model')
                 with gr.Row():
                     load_btn = gr.Button(value='Load model', elem_id='prompt_enhance_load', variant='secondary')
-                    load_btn.click(fn=self.load, inputs=[llm_model], outputs=[])
+                    load_btn.click(fn=self.load, inputs=[llm_model, use_openai], outputs=[])
                     unload_btn = gr.Button(value='Unload model', elem_id='prompt_enhance_unload', variant='secondary')
                     unload_btn.click(fn=self.unload, inputs=[], outputs=[])
                 with gr.Accordion('Custom model', open=False, elem_id='prompt_enhance_custom'):
@@ -938,7 +960,7 @@ class PromptEnhanceScript(scripts_manager.Script):
                         model_file = gr.Textbox(label='Model file', value=None, interactive=True, elem_id='prompt_enhance_model_file', placeholder='Optional GGUF model file inside GGUF model repo')
                     with gr.Row():
                         custom_btn = gr.Button(value='Load custom model', elem_id='prompt_enhance_custom_load', variant='secondary')
-                        custom_btn.click(fn=self.load, inputs=[model_repo, model_repo, model_gguf, model_type, model_file], outputs=[llm_model])
+                        custom_btn.click(fn=self.load, inputs=[model_repo, use_openai, model_repo, model_gguf, model_type, model_file], outputs=[llm_model])
                         llm_model.change(fn=self.get_custom, inputs=[llm_model], outputs=[model_repo, model_gguf, model_type, model_file])
                         gr.HTML('<br>')
                 with gr.Accordion('Options', open=False, elem_id='prompt_enhance_options'):
@@ -990,8 +1012,8 @@ class PromptEnhanceScript(scripts_manager.Script):
             # Update vision toggle interactivity when model changes
             llm_model.change(fn=self.update_vision_toggle, inputs=[llm_model], outputs=[use_vision], show_progress=False)
             if self.prompt:
-                apply_btn.click(fn=self.apply, inputs=[self.prompt, self.image, apply_prompt, llm_model, prompt_system, prompt_prefix, prompt_suffix, min_tokens, max_tokens, do_sample, temperature, repetition_penalty, top_k, top_p, thinking_mode, nsfw_mode, use_vision, prefill_text, keep_prefill, keep_thinking, custom_args, process_words, semantic_threshold, embedding_similarity], outputs=[prompt_output, self.prompt])
-        return [self.prompt, self.image, apply_auto, llm_model, prompt_system, prompt_prefix, prompt_suffix, min_tokens, max_tokens, do_sample, temperature, repetition_penalty, top_k, top_p, thinking_mode, nsfw_mode, use_vision, prefill_text, keep_prefill, keep_thinking, custom_args, process_words, semantic_threshold, embedding_similarity]
+                apply_btn.click(fn=self.apply, inputs=[self.prompt, self.image, apply_prompt, llm_model, prompt_system, prompt_prefix, prompt_suffix, min_tokens, max_tokens, do_sample, temperature, repetition_penalty, top_k, top_p, thinking_mode, nsfw_mode, use_vision, prefill_text, keep_prefill, keep_thinking, custom_args, process_words, semantic_threshold, embedding_similarity, use_openai], outputs=[prompt_output, self.prompt])
+        return [self.prompt, self.image, apply_auto, llm_model, prompt_system, prompt_prefix, prompt_suffix, min_tokens, max_tokens, do_sample, temperature, repetition_penalty, top_k, top_p, thinking_mode, nsfw_mode, use_vision, prefill_text, keep_prefill, keep_thinking, custom_args, process_words, semantic_threshold, embedding_similarity, use_openai]
 
     def after_component(self, component, **_kwargs): # searching for actual ui prompt components
         if getattr(component, 'elem_id', '') in ['txt2img_prompt', 'img2img_prompt', 'control_prompt', 'video_prompt']:
@@ -1002,7 +1024,7 @@ class PromptEnhanceScript(scripts_manager.Script):
             self.image.use_original = True
 
     def before_process(self, p: processing.StableDiffusionProcessing, *args, **kwargs): # pylint: disable=unused-argument
-        _self_prompt, self_image, apply_auto, llm_model, prompt_system, prompt_prefix, prompt_suffix, min_tokens, max_tokens, do_sample, temperature, repetition_penalty, top_k, top_p, thinking_mode, nsfw_mode, use_vision, prefill_text, keep_prefill, keep_thinking, custom_args, process_words, semantic_threshold, embedding_similarity = args
+        _self_prompt, self_image, apply_auto, llm_model, prompt_system, prompt_prefix, prompt_suffix, min_tokens, max_tokens, do_sample, temperature, repetition_penalty, top_k, top_p, thinking_mode, nsfw_mode, use_vision, prefill_text, keep_prefill, keep_thinking, custom_args, process_words, semantic_threshold, embedding_similarity, use_openai = args
         if not apply_auto and not p.enhance_prompt:
             return
         if shared.state.skipped or shared.state.interrupted:
@@ -1039,6 +1061,7 @@ class PromptEnhanceScript(scripts_manager.Script):
             process_words=process_words,
             semantic_threshold=semantic_threshold,
             embedding_similarity=embedding_similarity,
+            use_openai=use_openai,
         )
         timer.process.record('prompt')
         shared.state.end(jobid)