🎨 KoboldCpp Prompt Engine for ComfyUI

Transform simple ideas into high-production prompts. This node integrates local LLMs via KoboldCpp to expand basic input into descriptive Prompts, ranging from Danbooru tags to technical staging.

*I use it to creating complexed Wildcard*

UPDATE: Version 1.1 - The "Thinking" Update 🚀

Optimized for Reasoning Models (DeepSeek-R1, Qwen-Thinking, o1).

Key Features

• "Thinking" Mode: Enables Chain-of-Thought (CoT). The LLM plans the composition internally before generating the prompt.

• Automatic Filtering: Removes all internal reasoning (<think>...</think>) and meta-text (e.g., "Here is your prompt") so only the clean visual prompt reaches ComfyUI.

• Token Buffer: Automatically adds +250 tokens in Thinking mode to prevent prompts from being cut off by lengthy reasoning.

• Source Cleaner: Strips out dataset artifacts like ``.

How to Use

1. Select "Thinking" in the Mode dropdown.

2. Use a reasoning-capable model in KoboldCpp.

3. Note: start_helper is disabled in this mode to prioritize the <think> tag.

4. Enable "debug" to view the LLM's internal logic in the console.

Tip: If the output still truncates, increase max_tokens. Reasoning consumes a significant portion of the context window.

Workflow Components (Included)

This workflow automates prompt engineering by connecting your local LLM to ComfyUI. It requires the following custom nodes:

1. KoboldLLMPrompter: The core engine for prompt expansion.

2. Wildcard Saver: Automatically archives every generated prompt.

3. LazySmartWildcards: Manages dynamic inputs and wildcard processing.

🚀 Quick Start

    Installation: Save LLM_Wildcard.py in ComfyUI/custom_nodes/.

    Backend: Ensure KoboldCpp is running a compatible model (Llama, Mistral, or Qwen).

    Connection: Set the URL to your local API (default: http://127.0.0.1:5001).

🧠 Generation Modes:

SDXL (Tags) * Best For: SDXL / Pony-based models.   

• Output Style: Converts input into comma-separated Danbooru-style tags

Natural Sentence * Best For: Flux.1, SD3, or Midjourney-style prompting.

•     Output Style: Creates a cohesive, cinematic paragraph naturally fusing the subject, style, environment, and lighting.

Z-Engineer Best For: Qwen3-Z-Engineer Models* or similar high-parameter models.

•     Output Style: A production-focused, ~200-250 word paragraph with a deep focus on visual staging, lighting physics, and material textures.

🛠️ Key Functions:

    Style Selection:

• Choose from 14 aesthetics (e.g., Cyberpunk, DSLR, Anime) or use Random to cycle styles based on the seed.

    Start Helper:

• Force-starts the AI with specific phrases to bypass conversational "chatter" and ensure consistency.

    Filter:

• Internal logic that automatically strips AI artifacts like "Sure! Here is your prompt" and cleans up unfinished sentences.

⚙️ General Settings:

Temperature Advice:

• Use 0.2 – 0.5 for literal, prompt-loyal results.

• Use 0.8 – 1.2 for creative variety and unexpected descriptions.

Max Tokens Advice:

•         Low (50–150): Perfect for SDXL (Tags) to keep them punchy.

•         High (100): Necessary for Natural Sentence or 

•      High+ (250+): for Z-Engineer.

*I found this Model and the prompting template so effective that I decided to integrate them directly.


'VibeCoded' I'll try my best, or you can always ask Gemini or ChatGPT.