vLLM Prompt Node for ComfyUI

https://www.theoath.studio/projects/comfy-vllm-node?utm_source=civitai

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.

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

1. Restart ComfyUI.

Requirements

• A running vLLM server (see vLLM docs)

• Python package: requests (pip install requests)

• 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, restart your vLLM server. The node picks it up automatically.

Node Inputs

InputTypeDefaultDescriptionpromptSTRINGGeneration instruction. Supports {wild|card} syntax.prefixSTRINGmasterpiece, best quality, highresFixed tags prepended to the output. Not sent to the model.hostSTRINGlocalhostvLLM server host.portINT8765vLLM server port.max_tokensINT128Maximum tokens to generate.temperatureFLOAT0.7Sampling temperature. Higher = more creative.retriesINT3How many times to retry on empty or failed responses.

Node Output

OutputTypeDescriptioncombined_promptSTRINGprefix + generated text, ready to wire into CLIPTextEncode

The node displays a live preview after each generation showing:

• Prefix

• Raw generated text

• Final combined string

Wildcard Syntax

Use {option1|option2|option3} anywhere in your prompt. One option is chosen at random each run. Multiple wildcards are 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 forces the model to return comma-separated tags only:

### Stable Diffusion prompt tags (comma separated, no sentences):

Input: <your expanded prompt>

Output:

Generation stops at the first newline, preventing extra text.

If conversational output appears:

• Lower temperature to 0.3–0.5

• Use a larger model (≥ 1.5B recommended)

• Reduce max_tokens

Model Recommendations

ModelQualityNotesQwen2.5-0.5B⚠️ UnreliableToo small for consistent instruction followingQwen2.5-1.5B✓ UsableOccasional filler, mostly cleanQwen2.5-3B✓✓ RecommendedClean output, follows format reliablyQwen2.5-32B✓✓✓ BestOverkill but flawless

Tested With

• vLLM 0.4+

• Qwen2.5-0.5B, Qwen2.5-1.5B, Qwen2.5-3B

• ComfyUI (latest)