1. Workflow Overview

This workflow integrates DeepSeek-R1 32B LLM to automatically optimize AI art prompts, featuring:

• Transforming basic descriptions (e.g., "astronaut, dark sun, close-up") into professional-grade prompts

• Auto-adding lighting/material details and emotional tones

• Privacy-focused via Ollama local API

2. Core Components

3. Key Nodes

• Prompt Optimization Group

  • OllamaGenerateAdvance:

    • Input: Raw prompt (e.g., "astronaut, dark sun, close-up")

    • Parameters: temperature=0.9, top_p=0.8 (creativity control)

    • System Prompt: Requires 3-5 artistic elements + lighting/emotion details

  • TextSplitByDelimiter:

    • Splits output by <think>/</think> delimiters

• Output & Cleanup

  • ShowText|pysssss: Displays final optimized prompt

  • PurgeVRAM: Memory cleanup (critical for low-VRAM devices)

4. Dependencies

• Required Plugins:

  • ComfyUI-Ollama (local API connection)

  • pysssss (text display enhancement)

• Setup:

  • Local Ollama service (http://127.0.0.1:11434)

  • Model download: ollama pull deepseek-r1:32b

5. Input Parameters

• Mandatory:

  • Base prompt (e.g., "astronaut, dark sun, close-up")

• Tunable:

  • max_tokens=40 (output length control)

  • repeat_penalty=1 (reduces repetition)

6. Notes

⚠️ Requires 24GB+ VRAM (32B model size)
⚠️ Initial model loading takes 3-10 minutes
⚠️ Output includes <think> tags (remove manually)