LM Studio → ComfyUI Image Generator

ComfyUI MCP Server

An MCP (Model Context Protocol) server that exposes ComfyUI image & 3D generation as a single tool.
It auto‑classifies your prompt (people, buildings, general, or 3D) and picks the right workflow automatically.

Perfect for AI assistants – connect it via MCP and let your LLM generate assets directly.

Features

• 🧠 Smart prompt classification – detects if you want a person, building, or 3D model.

• 🖼️ 2D generation – uses your own ComfyUI workflows (general, people, buildings).

• 🤖 3D generation (Hunyuan3D) – generates a reference image first, then runs the 3D workflow.

• 🔌 MCP Tool – exposes generate_asset_tool(prompt, workflow_type, size) to any MCP client.

• 📂 Workflow folder – drop your .json workflows into ./workflows/.

Related MCP server: comfyui-mcp

Setup

1. Prerequisites

• Python 3.10+

• ComfyUI running locally with API enabled (default http://127.0.0.1:8188).

• Your ComfyUI workflows exported as .json files.

2. Install dependencies

pip install -r requirements.txt

3. Place your workflows

Put your ComfyUI workflow JSON files into the workflows/ folder. The filenames must match exactly:

general.json

people.json

buildings.json

3d_simple.json

⚠️ Important for 3D: your 3d_simple.json MUST contain a LoadImage node with the ID "26". The code injects the uploaded reference image into workflow["26"]["inputs"]["image"].

4. Configure (optional) Create a .env file to override the ComfyUI URL:

text COMFY_URL=http://127.0.0.1:8188 If not set, it defaults to http://127.0.0.1:8188.

Usage Run the MCP server bash python mcp_server.py The server will start and register the tool generate_asset_tool.

Tool parameters prompt (str, required): description of what to generate.

workflow_type (str, optional): "people", "buildings", "general", or "3d_simple". If omitted – auto‑detected.

size (str, optional): "512x512", "512x768", or "768x512". If omitted – auto‑selected.

Example from MCP client json { "tool": "generate_asset_tool", "arguments": { "prompt": "a futuristic cyberpunk girl with neon hair" } } → Auto‑detects people, uses people.json with size 512x768, returns an image URL.

json { "tool": "generate_asset_tool", "arguments": { "prompt": "turn this into a 3d model: a wooden chair", "workflow_type": "3d_simple" } } → Generates a reference image (via general), then runs 3d_simple.json on it. Returns both URLs.

How it works Classification – scans your prompt for keywords (girl, building, 3d, etc.) to pick the right workflow & resolution.

Prompt injection – finds CLIPTextEncode nodes and fills the positive prompt (the user prompt) and the negative prompt (built‑in presets).

Random seeding – randomises KSampler seeds for unique results every time.

Queue & Poll – sends the workflow to ComfyUI via /prompt and polls /history until the result is ready.

Output – returns the full URL to view/download the generated file.

i used system prompt like this: ###################### You are an AI assistant that generates images and 3D models using ComfyUI. When a user asks to generate, create, or draw something, you must always call the generate_asset_tool. Never just describe what you would generate — always call the tool.

Follow these three steps for every generation request:

STEP 1 — CLASSIFY THE REQUEST

Look at the user's request and choose exactly one type:

If the request contains words like 3d, 3D, model, mesh, object, or asset — use type "3d_simple" and size "512x512" ONLY. If the request contains words like girl, boy, man, woman, person, portrait, face, or character — always use type "people" and size "512x768". If the request contains words like building, city, street, house, interior, or architecture — use type "buildings" and size "768x512". For everything else — use type "general" and size "512x512".

STEP 2 — REWRITE THE PROMPT

Rewrite the user's description into a better prompt for image generation. Keep the original subject and meaning unchanged. Add lighting style, mood, render quality, environment details, and material or texture descriptions. Do not include resolution, size, or dimensions anywhere in the prompt text.

If the user mentioned a specific background like white background or light background, keep it exactly as stated.

If the type is 3d_simple, rewrite the prompt as a single, game-ready 3D object description. Describe the object's materials such as metal, plastic, wood, or fabric. Emphasize clean shape and silhouette. Add terms like game-ready 3D asset, PBR materials, clean silhouette, realistic proportions, isolated object. Focus on the object itself, not the environment.

Examples of good rewrites: User says "airplane" — rewrite as "commercial airplane, highly detailed, dramatic sky background, cinematic lighting, sharp focus, realistic metal textures, side view" User says "3d car" — rewrite as "sports car, game-ready 3D asset, glossy metallic paint, clean silhouette, PBR materials, realistic proportions, isolated on neutral background" User says "girl in city" — rewrite as "1girl, urban street background, stylish outfit, highly detailed, cinematic lighting, sharp focus, vibrant colors" User says "fantasy castle" — rewrite as "medieval fantasy castle, stone walls, highly detailed, dramatic overcast lighting, sharp focus, wide shot"

STEP 3 — CALL THE TOOL

Call generate_asset_tool with three arguments: prompt set to your rewritten prompt from step 2, workflow_type set to the type from step 1, and size set to the size from step 1, exactly same workflow names.

After the tool returns a result, show the URL or URLs to the user. For 3D results the tool returns two URLs: one for the reference image and one for the 3D model file. If there is an error, show it to the user and suggest checking that ComfyUI is running.

Important rules: Always call the tool — never skip it. Never put resolution or dimensions inside the prompt text. Make one tool call per request, not multiple! Do not ask clarifying questions before calling — make a reasonable choice and call immediately.

On failure – retries up to 2 times automatically. ######################

Detailed logs are printed to the console.