How the Negative Prompt Is Passed to the Generation API in Open-Generative-AI

In the Open-Generative-AI application, the negative prompt flows from the UI to the local diffusion model via the -n CLI flag, but is currently discarded when using the remote Muapi cloud service.

The Anil-matcha/Open-Generative-AI repository provides a unified desktop interface for generative AI models, supporting both local inference and cloud-based generation. Understanding how the negative prompt is passed to the generation API requires examining two distinct execution paths: a fully implemented local pipeline that translates the parameter into command-line arguments, and a remote integration where the field is documented but omitted from the actual HTTP payload.

Local Inference Path (Desktop/Electron)

When running models locally, the application captures the negative prompt from the user interface and propagates it through three distinct layers before reaching the Stable Diffusion binary.

UI Capture in ImageStudio.js

The negative prompt originates in the advanced settings panel of the image generation studio. In src/components/ImageStudio.js, a local variable tracks the input value through an event listener:

// src/components/ImageStudio.js (lines 44-46, 600)
let negativePrompt = '';
const negPromptInput = advancedPanel.querySelector('#negative-prompt-input');
if (negPromptInput) {
    negPromptInput.oninput = (e) => {
        negativePrompt = e.target.value;
    };
}

Generation Request Construction

When the user initiates generation, the negativePrompt variable is passed to the local AI bridge along with other parameters. At line 1192 of ImageStudio.js, the code explicitly includes negative_prompt in the options object:

// src/components/ImageStudio.js (line 1192)
const res = await localAI.generate({
    model: selectedLocalModel,
    prompt,
    negative_prompt: negativePrompt || undefined,
    aspect_ratio: selectedAr,
    steps,
    guidance_scale: guidanceScale,
    seed,
});

Electron CLI Translation

The localAI.generate() method forwards the parameters to the Electron main process. Inside electron/lib/localInference.js, the code detects the presence of negative_prompt and appends the -n flag to the arguments array passed to the sd-cli binary:

// electron/lib/localInference.js (lines 84-86)
if (params.negative_prompt) {
    args.push('-n', params.negative_prompt);
}

The local inference layer effectively translates the JavaScript property into a command-line argument that the underlying Stable Diffusion CLI understands, ensuring the model steers away from the unwanted concepts specified by the user.

Remote API Path (Muapi Cloud)

The cloud generation path through Muapi presents a different scenario where the negative prompt parameter is declared but not transmitted.

Declared But Not Implemented

In src/lib/muapi.js, the JSDoc for the generateImage method explicitly documents a negative_prompt parameter at line 20:

// src/lib/muapi.js (line 20)
/**
 * @param {string} params.negative_prompt
 */
async generateImage(params) { /* … */ }

However, the actual payload construction (lines 35-40) omits this field entirely. The finalPayload object sent via POST request only includes prompt, aspect_ratio, resolution, and other select fields, while the negative_prompt value is discarded before the network request is dispatched.

This creates a functional asymmetry between the two modes: local inference fully supports negative prompting, while the remote Muapi integration acts as a placeholder awaiting backend support.

Technical Implementation Summary

The codebase reveals a clear architectural distinction in how parameters flow through the system:

Component	File Path	Role in Negative Prompt Handling
UI State	src/components/ImageStudio.js	Captures input from #negative-prompt-input and stores it in the negativePrompt variable.
Local Bridge	electron/lib/localInference.js	Translates the negative_prompt property into the -n CLI flag for the sd-cli binary.
Remote Client	src/lib/muapi.js	Documents the parameter in JSDoc but excludes it from the POST payload (lines 35-40).

Summary

• Local mode propagates the negative prompt from the React UI (ImageStudio.js) through the Electron IPC layer (localInference.js), ultimately converting it to a -n command-line argument for the local Stable Diffusion binary.
• Remote mode currently drops the negative prompt; while the API client interface in muapi.js accepts the parameter, the implementation does not include it in the HTTP request to the Muapi endpoint.
• The negative_prompt field is optional in both paths, defaulting to undefined when empty to prevent passing empty strings to the CLI or API.

Frequently Asked Questions

Does the remote Muapi endpoint support negative prompts?

Currently, no. While the generateImage method in src/lib/muapi.js documents a negative_prompt parameter in its JSDoc signature at line 20, the actual payload construction (lines 35-40) does not include this field in the POST request. This is a known limitation in the current codebase that prevents negative prompts from reaching the cloud service.

What CLI flag does the local inference use for negative prompts?

The Electron-based local inference translates the negative_prompt property into the -n flag. In electron/lib/localInference.js at lines 84-86, the code checks if (params.negative_prompt) and executes args.push('-n', params.negative_prompt) to pass the value to the sd-cli binary.

Where is the negative prompt stored in the application state?

The negative prompt is stored in a module-level variable named negativePrompt declared at line 44 of src/components/ImageStudio.js. This variable is updated via an oninput event listener attached to the #negative-prompt-input DOM element at line 600, and is later referenced when constructing the generation request at line 1192.

Is negative prompt support planned for the remote API?

The codebase structure suggests placeholder support has been prepared, as evidenced by the JSDoc annotation in src/lib/muapi.js. However, the actual implementation requires backend support from the Muapi service and additional client-side code to include the parameter in the payload construction logic before the feature becomes functional.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}