Description
I am reporting a significant language-mixing issue when using the Gemma-4-26B Vision Language Model (VLM) within this suite. Despite explicitly setting both the custom_prompt and system_prompt to require Chinese-only output, the model continues to generate large amounts of English text.
Through code analysis of nodes.py, I have identified that the root cause is the hardcoded English strings within the preset_prompts dictionary. These English instructions are prepended to the user's prompt, creating a "language bias" that causes global models like Gemma-4 to default to English or use English for "Chain of Thought" reasoning.
Technical Evidence
- Source Code Analysis (from nodes.py)
The plugin defines several presets in the nodes.py file that are entirely in English:
"Normal - Describe": "Describe this @."
"Prompt Style - Extreme Detailed": "Generate an extremely detailed and descriptive text-to-@ prompt from the @..."
In the process function, the plugin concatenates these English strings with the user's prompt:
Python
p = preset_prompts[preset_prompt].replace("#", custom_prompt.strip()).replace("@", "video" if video_input else "image")
user_content.append({"type": "text", "text": p})
This results in the model receiving a prompt like: "Describe this image. [User's Chinese Prompt]". For Gemma-4-26B, the initial English instruction sets a high-priority "language context," leading to English output.
- Visual Evidence (Verified from Screenshots)
Gemma-4-26B (Failure Case): As seen in image_cd9f21.jpg and image_cdfa10.jpg, the model loader is explicitly using gemma-4-26B\gemma-4-26B-A4B-it-UD-IQ4_XS.gguf.
The Result: The output in the "Show Text" node contains an English <|channel|>thought block and English headers (e.g., "Main Focus", "Desk Items", "Background"), directly following the structure of the English preset instead of the Chinese instruction.
Steps to Reproduce
Load Gemma-4-26B (GGUF) using the Llama-cpp Model Loader.
Set preset_prompt to "Normal - Describe" or "Extreme Detailed".
Input a Chinese request in custom_prompt (e.g., "请详细描述图片内容").
Observe the output containing English headers, reasoning, or full English descriptions.
Suggested Fix
Localization of Presets: Allow the preset_prompts in nodes.py to be localized (Chinese version) or detect the user's input language.
User Control: When a custom_prompt is provided, the English preset_prompt should not be prepended unless the user specifically toggles a "Combine with Preset" option.
Prompt Weighting: Ensure that system instructions regarding language (e.g., "Respond strictly in Chinese") are placed after any preset templates to ensure they carry higher weight in the model's attention mechanism.
Attachments
image_cd9f21.jpg / image_cdfa10.jpg: Showing Gemma-4-26B generating English thought-chains and headers despite Chinese input.
Description
I am reporting a significant language-mixing issue when using the Gemma-4-26B Vision Language Model (VLM) within this suite. Despite explicitly setting both the custom_prompt and system_prompt to require Chinese-only output, the model continues to generate large amounts of English text.
Through code analysis of nodes.py, I have identified that the root cause is the hardcoded English strings within the preset_prompts dictionary. These English instructions are prepended to the user's prompt, creating a "language bias" that causes global models like Gemma-4 to default to English or use English for "Chain of Thought" reasoning.
Technical Evidence
The plugin defines several presets in the nodes.py file that are entirely in English:
"Normal - Describe": "Describe this @."
"Prompt Style - Extreme Detailed": "Generate an extremely detailed and descriptive text-to-@ prompt from the @..."
In the process function, the plugin concatenates these English strings with the user's prompt:
Python
p = preset_prompts[preset_prompt].replace("#", custom_prompt.strip()).replace("@", "video" if video_input else "image")
user_content.append({"type": "text", "text": p})
This results in the model receiving a prompt like: "Describe this image. [User's Chinese Prompt]". For Gemma-4-26B, the initial English instruction sets a high-priority "language context," leading to English output.
Gemma-4-26B (Failure Case): As seen in image_cd9f21.jpg and image_cdfa10.jpg, the model loader is explicitly using gemma-4-26B\gemma-4-26B-A4B-it-UD-IQ4_XS.gguf.
The Result: The output in the "Show Text" node contains an English <|channel|>thought block and English headers (e.g., "Main Focus", "Desk Items", "Background"), directly following the structure of the English preset instead of the Chinese instruction.
Steps to Reproduce
Load Gemma-4-26B (GGUF) using the Llama-cpp Model Loader.
Set preset_prompt to "Normal - Describe" or "Extreme Detailed".
Input a Chinese request in custom_prompt (e.g., "请详细描述图片内容").
Observe the output containing English headers, reasoning, or full English descriptions.
Suggested Fix
Localization of Presets: Allow the preset_prompts in nodes.py to be localized (Chinese version) or detect the user's input language.
User Control: When a custom_prompt is provided, the English preset_prompt should not be prepended unless the user specifically toggles a "Combine with Preset" option.
Prompt Weighting: Ensure that system instructions regarding language (e.g., "Respond strictly in Chinese") are placed after any preset templates to ensure they carry higher weight in the model's attention mechanism.
Attachments
image_cd9f21.jpg / image_cdfa10.jpg: Showing Gemma-4-26B generating English thought-chains and headers despite Chinese input.