<h1 id="vllm-prompt-node-for-comfyui">vLLM Prompt Node for ComfyUI</h1><p><a target="_blank" rel="ugc" href="comfy-vllm-node">https://www.theoath.studio/projects/comfy-vllm-node?utm_source=civitai</a><br /></p><p>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.<br /></p><h1 id="installation">Installation</h1><ol><li><p>Clone or copy this folder into your <code>ComfyUI/custom_nodes/</code> directory:</p></li></ol><pre><code>cd ComfyUI/custom_nodes
git clone https://github.com/OATH-Studio/comfy-vLLM</code></pre><ol><li><p>Restart ComfyUI.</p></li></ol><hr /><h2 id="requirements">Requirements</h2><ul><li><p>A running vLLM server (see <a target="_new" rel="ugc" href="https://docs.vllm.ai">vLLM docs</a>)</p></li><li><p>Python package: <code>requests</code> (<code>pip install requests</code>)</p></li><li><p>ComfyUI</p></li></ul><hr /><h2 id="setup">Setup</h2><p>Start your local vLLM server. The node will automatically detect whichever model is currently loaded. No need to specify it in the node.</p><p>Example launch:</p><pre><code>vllm serve ./models/Qwen2.5-3B \
--host 0.0.0.0 \
--port 8765 \
--served-model-name Qwen2.5-3B</code></pre><p>Note: The node queries <code>/v1/models</code> on each generation and uses the first model returned. If you change models, restart your vLLM server. The node picks it up automatically.</p><hr /><h2 id="node-inputs">Node Inputs</h2><p>InputTypeDefaultDescriptionpromptSTRINGGeneration instruction. Supports <code>{wild|card}</code> 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.</p><hr /><h2 id="node-output">Node Output</h2><p>OutputTypeDescriptioncombined_promptSTRINGprefix + generated text, ready to wire into <code>CLIPTextEncode</code></p><p>The node displays a live preview after each generation showing:</p><ul><li><p>Prefix</p></li><li><p>Raw generated text</p></li><li><p>Final combined string</p></li></ul><hr /><h2 id="wildcard-syntax">Wildcard Syntax</h2><p>Use <code>{option1|option2|option3}</code> anywhere in your prompt. One option is chosen at random each run. Multiple wildcards are resolved independently.</p><pre><code>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}</code></pre><p>Wildcards are expanded before the prompt is sent to the model, so the model always receives a fully resolved string.</p><hr /><h2 id="example-workflow">Example Workflow</h2><pre><code>VLLMPromptNode ──→ CLIPTextEncode (positive) ──→ KSampler
       ↑
CLIPTextEncode (negative) ───┘</code></pre><hr /><h2 id="prompt-format">Prompt Format</h2><p>The node uses the completions endpoint with a structured format that forces the model to return comma-separated tags only:</p><pre><code>### Stable Diffusion prompt tags (comma separated, no sentences):

Input: &lt;your expanded prompt&gt;

Output:</code></pre><p>Generation stops at the first newline, preventing extra text.</p><p>If conversational output appears:</p><ul><li><p>Lower temperature to <code>0.3–0.5</code></p></li><li><p>Use a larger model (≥ 1.5B recommended)</p></li><li><p>Reduce <code>max_tokens</code></p></li></ul><hr /><h2 id="model-recommendations">Model Recommendations</h2><p>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</p><hr /><h2 id="tested-with">Tested With</h2><ul><li><p>vLLM 0.4+</p></li><li><p>Qwen2.5-0.5B, Qwen2.5-1.5B, Qwen2.5-3B</p></li><li><p>ComfyUI (latest)</p></li></ul><p></p>

Files with SHA256: c514aa7224c8b6edeb31fb08c4f19f21c4694ae053bec3e004c5aefb9b9a867c