Getting started with images

What Easy Setup installs when you say yes to images, how big it is, what the first picture looks like, and what to do when the picture comes out black. If you have not installed LumaBrowser yet, start with the main guide; this page picks up at the image toggle.

What you need

A graphics card with 8 GB of video memory or more. NVIDIA (GeForce RTX 3060 12 GB and up is comfortable), AMD (RX 6700 and up), or an Apple Silicon Mac with 16 GB or more.
About 10 GB of free disk for the engine and the first model, more if you add others.
No Python, no separate installer, no driver setup. The app fetches the right engine build for your card (CUDA for NVIDIA, Vulkan for AMD and Intel, native on Apple Silicon).

No graphics card? Images will render on the processor, but a single picture takes several minutes. It works, and it is a fine way to see one cat, but it is not recommended for real use. Chat and stories are unaffected.

What Easy Setup picks

Easy Setup chooses one image model for your card and installs it next to the chat model:

Your card	What it picks	Download
8 to 12 GB	Z-Image Turbo, or SDXL 1.0 on older cards	about 6 GB
12 to 24 GB	Z-Image Turbo or FLUX.1 schnell	6 to 12 GB, larger FLUX variants up to about 30 GB
24 GB and up, or a 32 GB+ Mac	FLUX.1 schnell plus the Qwen-Image-Edit editing model	12 to 30 GB, plus about 18 GB for editing

The plan screen shows the exact size before anything downloads. If the total looks like too much for your disk or connection, untick the editing model; it can be added later.

Screenshot placeholder: the Easy Setup plan screen with both a chat model line and an image model line ("Images: FLUX.1 schnell, about 12 GB").

The plan with images on. Two downloads, one Start button.

The first picture

When setup finishes, the chat opens. Type make an image of a tabby cat asleep in a sunbeam.
The chat shows a "generating" card. The first render takes about a minute because the model has to load into the card; on a mid-range card later renders take a few seconds, on a high-end card about a second.
The picture appears in the conversation. Click it to open it full size, save it, or ask for changes ("make it night time", "same cat, now on a windowsill").

Screenshot placeholder: the chat with a generated image of a cat in a sunbeam inside an image card, and a follow-up request below it.

The first render. Ask for changes in the same conversation.

Want more control? The Image tab has the full panel: size, steps, seed, reference images for keeping the same face across pictures, painting over a region, and the model catalog. Descriptions of each model and what it is good at are on the Create page.

Adding images after the first run

If you said no to images in the wizard, open the Image tab and click Easy Setup. It runs the same plan for just the image half. In a future release, asking for an image in chat when images are not installed will offer this with one click; for now the Image tab is the door.

Common failures and fixes

The picture is black, or a flat grey square

The card ran out of memory partway through, usually because the resolution or the model is too big for it. Fixes, in order: lower the size in the Image tab (1024 by 1024 down to 768 by 768), pick a smaller model (Z-Image Turbo or SDXL instead of FLUX), close a game or browser tab that is also using the card, and if you keep both a generation and an editing model loaded, unload the one you are not using.

"Out of memory" in the generating card

Same cause, reported before the render started. Same fixes. On cards with 8 GB, stick with Z-Image Turbo, SDXL, or SD 1.5.

"Cannot start: server is ready"

A known bug in 1.8.1: the Image tab's Easy Setup refuses to run while the chat engine is already running. Workaround: open the Setup tab, stop the chat server, run the image Easy Setup, then start chat again. The fix is in progress.

The download stalls at 90 percent

Downloads resume. Pause it and resume, or restart the app. If your disk is full, free space or move the models folder to another drive from Settings, then resume.

It is very slow on my NVIDIA card

Check the Image tab's runtime line. If it says CPU or Vulkan on an NVIDIA card on Windows, click Easy Setup again to fetch the CUDA build. On Linux, Vulkan is currently the only option for NVIDIA; see below.

The picture does not match what I asked for

Say more, and say it plainly: subject, setting, lighting, style. Small models follow short, concrete descriptions best. The generating card shows the exact prompt the model received, so you can see what got lost.

Linux, honestly

In 1.8.1 the automatic image install does not work on Linux. Two problems, both ours: the installer cannot unpack the zip it downloads, and on NVIDIA it picks a CUDA build that has no Linux version. Chat is unaffected.

Workaround until the fix ships: download the Linux Vulkan (or CPU) build of stable-diffusion.cpp from its GitHub releases, unzip it anywhere, then in the Image tab choose Locate runtime and point at the folder. Pick a model from the catalog after that; model downloads work. A rendered picture on a laptop processor takes about three minutes; on a Vulkan-capable card a few seconds.

A note on content

The image models and roleplay mode apply no content filter of their own. What they can depict depends on the model you load, and they are intended for adults. You are responsible for the models you choose and what you generate on your own machine.

Characters that stay consistent from scene to scene: Roleplay. Talk instead of type: Voice mode. Describe a game and play it: Small games. Every model and what it needs: the Create page.