How to use AI locally on a PC?

One of the most exciting aspects of delving into a new technology is getting hands-on and truly immersing oneself in the experience. In the past, we’ve explored LLMs and self-trained models for feature extraction, along with pre-trained models. A recurring theme in every 2023 blog post is the automatic creation of images using Midjourney or other AI tools. However, these tools typically operate within a cloud system. The resulting images can range from realistic to artistic, depending on the model employed.

How to start?

Finding a starting point is no easy feat. So far, I’ve tested Automatic1111, ComfyUI, Gradio and InvokeAI as user interfaces. Each has its pros and cons, even varying in computational speed.

For me, InvokeAI stands out as the simplest and most user-friendly solution currently available. The installation is quick and straightforward, and the user-friendliness holds up well under different conditions.

1.Requirements for using InvokeAI

InvokeAI is compatible with Linux, Windows, and Mac. Linux users have the option to utilize Nvidia-based cards (with CUDA support) or AMD cards (with ROCm driver).

Automatic Installer: Ideal for Beginners

This script takes care of installing all the necessary third-party libraries and InvokeAI itself. It includes a “Developer Console” for debugging and accessing experimental features. Our focus here will be on the automatic installer.

Manual Installation: For Experienced Users and Developers

Refer to InvokeAI’s installation page: https://invoke-ai.github.io/InvokeAI/installation/INSTALLATION/#manual-installation

System Requirements

Your system should have one of the following:

• a Nvidia-based graphics card with at least 4 GB VRAM OR
• an AMD-based graphics card with at least 4 GB VRAM (Linux only) OR
• an Apple computer with an M1 chip

SDXL 1.0 Specifications & Requirements:

• a Nvidia-based graphics card with 8 GB or more VRAM
• an AMD-based graphics card with 16 GB or more VRAM (Linux only)
• an Apple computer with an M1 chip

Memory and Storage

Minimum requirements are:

• at least 12 GB of RAM (Random Access Memory)
• a minimum of 18 GB of free disk space for the machine learning model, Python, and its dependencies.

2. Installation

10 Steps Guide for the Automatic Installer (Ideal for First-time Users)

1. Navigate to the end of the Latest Release Page.
2. Select and download the .zip file that corresponds to your operating system (Windows, macOS, or Linux).
3. After downloading, extract the file.
4. For Windows users: Double-click on the install.bat script. For macOS: Open a terminal, drag the install.sh file from the Finder to the terminal, and press Enter. For Linux: Execute the install.sh script.
5. During the installation, you’ll be prompted to choose a directory for InvokeAI and its image files for image generation. Choose a location with at least 15 GB of free disk space, more if you plan to add additional models.
6. Wait patiently while the installer sets up everything. After installation, a script will launch to configure InvokeAI and select initial models for image generation.
7. Locate the installation directory of InvokeAI (note that this is not the same as where you unpacked the zip file!). By default, this is ~/invokeai on Linux/Mac and C:\Users\YourName\invokeai on Windows. In this directory, you’ll find startup scripts named invoke.sh and invoke.bat.
8. To start on Windows, double-click on the invoke.bat file. On macOS and Linux, open a terminal, drag the invoke.sh file into the terminal, and press Enter.
9. Once the software is running, press 2 to open the “browser-based user interface,” press Enter, and give Stable Diffusion a few minutes to start.
10. Then, in your web browser, navigate to http://localhost:9090.

3. The Models

Using the right models is crucial to your success in image generation. The easiest approach is to find a so-called “pre-trained” model, where the additional learning steps in SDXL are already integrated. For our example here, I suggest using the Realvis 3.0 SDXL Baked from https://civitai.com/. This model is trained for photorealism and will deliver excellent results if you use the right prompts. Another secret for great results is the so-called “Refiner.” This additional step adds the final 10% of quality to your results. I’m using the “stable-diffusion-xl-refiner-1-0″.

Installing this model is quite straightforward; you just need to paste the link into the Import Models dialog in InvokeAI.

4. Runtime and the fun begins

Let’s have some fun and use some real good prompts.

Prompt:

Hyper-realistic full-body image of a captivating female cybernetic android that will overtake the world with a wide, genuine smile and deep, piercing blue eyes. She boasts a bold blond buzzcut hairstyle and sophisticated dark makeup, highlighting her perfect skin. The image is taken in white plastic clothes and in a futuristic spaceship with warm, soft light, capturing her face and shoulders and emphasizing fine details like the texture of her skin and the subtle shine in her eyes. The photograph, styled as a high-fashion portrait, conveys a blend of natural beauty and contemporary style, ideal for a high-quality magazine cover.

Style:

comic style, 8k, agfa vista, photographic, high quality, realistic face

Negative Prompt:

(worst quality, low resolution, minimal details, oversaturated or undersaturated colors, overexposed or underexposed lighting, grayscale, black and white, poor photography techniques:1.4), (any watermarks, signatures, text fonts, usernames, errors, logos, words, letters, digits, autographs, trademarks, names:1.2), (blurry or grainy textures), grotesque, unsightly features, asymmetry, malformed or mutilated elements, poorly lit areas, bad shadowing, drafts, cropped or out of frame subjects, censored parts, jpeg artifacts, out of focus areas, glitches, duplicates, (airbrushed, cartoonish, anime-style, semi-realistic, CGI, rendered, digital art, manga, amateurish effects:1.3), (3D models, 3D game scenes, 3D characters:1.1), (incorrect anatomy, disproportionate body parts, facial distortions, imperfect teeth, awkward limbs:1.3)

Use the DPM++ SDE Karras sampler, 4-10 steps, and CFG Scale 1-2.5.

Final Conclusion

Generating images on a local computer is no longer a challenge these days. It’s quite easily accessible and reproducible on almost any computer with a powerful graphics card. In essence, it would work on any other computer, but unfortunately, the wait times would be considerably long. Thus, it’s only enjoyable on systems where there’s a certain level of hardware support.

Frequently ask questions: