vLLM Prompt Node for ComfyUI

A ComfyUI custom node that generates Stable Diffusion prompts using a locally running vLLM server. Supports wildcard expansion and a fixed prefix for quality tags or style anchors.

https://www.theoath.studio/projects/comfy-vllm-node

Requirements

• A running vLLM server (see vLLM docs)
• Python package: requests (pip install requests)
• ComfyUI

Installation

1. Clone or copy this folder into your ComfyUI/custom_nodes/ directory:

cd ComfyUI/custom_nodes
git clone https://github.com/OATH-Studio/comfy-vLLM

2. Restart ComfyUI.

Setup

Start your local vLLM server. The node will automatically detect whichever model is currently loaded no need to specify it in the node.

Example launch:

vllm serve ./models/Qwen2.5-3B \
    --host 0.0.0.0 \
    --port 8765 \
    --served-model-name Qwen2.5-3B

Note: The node queries /v1/models on each generation and uses the first model returned. If you change models, just restart your vLLM server the node picks it up automatically.

Node Inputs

| Input | Type | Default | Description | |---|---|---|---| | prompt | STRING | | Generation instruction. Supports {wild\|card} syntax. | | prefix | STRING | masterpiece, best quality, highres | Fixed tags prepended to the output. Not sent to the model. | | host | STRING | localhost | vLLM server host. | | port | INT | 8765 | vLLM server port. | | max_tokens | INT | 128 | Maximum tokens to generate. | | temperature | FLOAT | 0.7 | Sampling temperature. Higher = more creative. | | retries | INT | 3 | How many times to retry on empty or failed responses. |

Node Output

| Output | Type | Description | |---|---|---| | combined_prompt | STRING | prefix + generated text, ready to wire into CLIPTextEncode |

The node displays a live preview on the node face after each generation showing the prefix, the raw generated text, and the final combined string separately so you can see exactly what was assembled.

Wildcard Syntax

Use {option1|option2|option3} anywhere in your prompt. One option is chosen at random each run. Multiple wildcards are each resolved independently.

A {red|blue|green} dragon, {breathing fire into the sky|coiled around a mountain peak in a storm|diving into a glowing ocean abyss|rearing up against a blood moon}

Wildcards are expanded before the prompt is sent to the model, so the model always receives a fully resolved string.

Example Workflow

VLLMPromptNode ──→ CLIPTextEncode (positive) ──→ KSampler
                                                      ↑
                         CLIPTextEncode (negative) ───┘

Prompt Format

The node uses the completions endpoint with a structured format that instructs the model to return comma-separated tags only, with no conversational filler:

### Stable Diffusion prompt tags (comma separated, no sentences):
Input: <your expanded prompt>
Output:

Generation stops at the first newline, so the model cannot ramble past a single line of tags. If you still get conversational responses, try:

• Lowering temperature to 0.3–0.5
• Increasing to a larger model (1.5B or 3B recommended minimum)
• Reducing max_tokens to leave less room for filler

Model Recommendations

| Model | Quality | Notes | |---|---|---| | Qwen2.5-0.5B | ⚠️ Unreliable | Too small for consistent instruction following | | Qwen2.5-1.5B | ✓ Usable | Occasional filler, mostly clean | | Qwen2.5-3B | ✓✓ Recommended | Clean output, follows format reliably | | Qwen2.5-32B | ✓✓✓ Best | Overkill for this task but flawless |

Tested With

• vLLM 0.4+
• Qwen2.5-0.5B, Qwen2.5-1.5B, Qwen2.5-3B
• ComfyUI (latest)