---
title: "Create with an AI Agent — How to Use Creative AI Models Locally"
canonical: "https://build.nvidia.com/playbooks/agents-for-creators/local-creative-models.md"
---

# How to Use Creative AI Models Locally

The way to run creative AI models locally is through ComfyUI. It’s a powerful tool that automates the download of the model, the installation, and running the model. You can generate everything from 3D, to video to image with the ComfyUI App, or simply ask your agent to control it.

## Step 1: Install ComfyUI.

[Download Comfy Desktop](https://comfy.org/download) and follow its Windows installation guide for your device.

Once installed, open Comfy Desktop and select the **ComfyUI** instance to launch the app.

![Comfy Desktop launcher with the ComfyUI instance card highlighted](https://assets.ngc.nvidia.com/products/api-catalog/rtx-spark/comfy-desktop-instance.png)

You have 2 options on how to use ComfyUI:

- Directly in the ComfyUI App
- Through an Agent

The steps below walk you through both.

## Step 2: Generate assets in the ComfyUI App.

1. Open the Template browser in the left sidebar, under **Templates**.

![ComfyUI left sidebar with the Templates icon selected](https://assets.ngc.nvidia.com/products/api-catalog/rtx-spark/templates-button-location.png)

2. Select the workflow you are interested in. We recommend starting with something simple, like **FLUX.2 [Klein] 4B: Text to Image**. This template is also the best starting point on a 32 GB RTX Spark.

3. Errors will pop up since the workflow needs to download its required models. Click on **Details**, then **Download all models**. Wait for all the models to download. You can monitor progress using the Downloads icon at the top.

> [!TIP]
> Some models come in smaller NVFP4 (4-bit) versions that run faster and use less memory. On a 32 GB RTX Spark, use the NVFP4 version of FLUX.2 Klein 4B so ComfyUI and your agent fit in memory together. Your agent can find and set it up for you. See Step 3 below.

4. Run the workflow. It will generate an asset from the provided text prompt. Feel free to adjust the initial prompt that is in the first node.

5. You can find your generated images in Assets on the left sidebar. Save the output image using the Download icon.

![ComfyUI Assets panel showing a generated hedgehog image and its Download icon](https://assets.ngc.nvidia.com/products/api-catalog/rtx-spark/comfyui-generated-hedgehog.png)

## Step 3: Use your agent to run ComfyUI.

1. Open your Agent and connect it to ComfyUI. ComfyUI offers an official MCP (a way to connect your agent to different apps), it’s simple, run this prompt in your agent:

```prompt
Install and connect the official ComfyUI MCP server from Comfy-Org to Hermes on my Windows ARM64 RTX Spark. Comfy Desktop is already installed, so reuse what’s working and install only what’s missing, keeping my other connections unchanged. Guide me only through steps I need to do myself, and troubleshoot only if something fails. Finish by confirming the connection works and showing five example workflow templates—no need to check every template or download models.
```

2. Switch to a faster model. Your agent can pick a model on its own when you describe what you want, but for the best speed on RTX Spark we recommend telling it which one to use.

For FLUX.2 Klein 4B, ask for the official [NVFP4 variant](https://huggingface.co/black-forest-labs/FLUX.2-klein-4b-nvfp4). NVFP4 is NVIDIA’s 4-bit floating-point format. These variants are smaller than the full-precision models and can run faster on NVIDIA RTX 50 Series GPUs and RTX Spark with compatible software. Ask your agent to check that your installed ComfyUI supports the variant before downloading it.

> [!NOTE]
> On a 32 GB RTX Spark, do this step before you start generating with your agent. The NVFP4 version leaves room for both ComfyUI and your agent’s model.

```prompt
Find the official Black Forest Labs FLUX.2 Klein 4B NVFP4 model and check that it works with my installed ComfyUI. Download any missing required files and configure a compatible text-to-image workflow to use it. Preserve my existing workflows. If this configuration is unsupported, explain what is missing before choosing another variant.
```
3. Now you can start generating assets! You can simply request the agent for what you want and it will create the prompt and adjust the settings for you. You can also load a folder into the prompt and request where you want the assets to be saved.