How to run an AI image generator on your own PC

Online image generators charge in credits, keep your prompts and upload every photo you add. The same kind of picture can be made on your own PC, privately and without a meter running, if your graphics card is up to it. Here is what that takes and how to get good results.

Why the graphics card matters most

An image model is a large file of learned numbers, and making a picture means running through it many times. Graphics cards do that work far faster than a processor, but only if the model fits in the card's own memory, called video memory or VRAM. That number, not the card's name, decides what you can run.

• Find your VRAM. Open Task Manager, choose Performance, then your GPU. Dedicated GPU memory is the figure that counts; shared memory borrowed from the PC is much slower.
• Smaller, older models such as SD 1.5 run on modest cards with a few gigabytes.
• Larger models such as SDXL and newer designs give better detail and follow prompts more closely, and need more memory, typically 6 GB or more for comfortable use.
• Integrated graphics on most laptops can run small models only slowly, if at all.

Memory, storage and patience

Models are several gigabytes each, so plan for plenty of free space on a fast SSD, and 16 GB of system memory or more. The first picture after starting takes longest while the model loads; after that each picture takes seconds to a minute depending on the card, the size and how many steps you ask for.

Why the face keeps changing

Each picture starts from random noise, so the same prompt gives a different person every time unless something holds the face steady. The usual workarounds are swapping a face onto the finished picture afterwards, which often looks pasted on, or training a small model on one person. Newer approaches let you save a person from a few photos and carry that face into new pictures as they are made, which looks more natural and takes minutes instead of hours.

Getting better pictures

1. Describe the scene, not just the subject. Say who, doing what, where, in what light and from what camera angle. Short prompts leave the model to guess.
2. Make several variations. Ask for four or more at once and keep the best, rather than rewriting the prompt after every miss.
3. Edit instead of starting again. If one detail is wrong, paint over that area and redraw only it.
4. Upscale at the end. Make the picture at the model's natural size, then enlarge it with an upscaler that rebuilds detail.

Using AI images responsibly

• Consent. Use photos of other people only with their permission, and never to mislead.
• Labelling. Say a picture is AI generated when you share it, especially if it looks like a real photo.
• Model licences. Model files you download come with their own licences, and some limit commercial use. Check before you sell work made with them.
• Check the result. AI pictures can get hands, text and details wrong. Look closely before you publish.

How XLUNA AI Fusion Ultra handles this

XLUNA AI Fusion Ultra makes pictures on your own graphics card, so your prompts and photos never leave your PC and nothing is billed per picture. With 6 GB of video memory or more it runs every model, including the drawing model that comes with the app; with less it runs SD 1.5 and SDXL models you add. Personas keep one face in every picture from up to three photos, with no face swap, and Image Fusion edits a photo from a plain sentence. A content filter that cannot be turned off is always on, and every result is labelled as AI generated.