Generate images on your Mac.

1Image models and the memory they need

The makers' own memory figures are for graphics cards at full size: FLUX.2 Klein 4B runs in about 13 GB of video memory and Z-Image Turbo within 16 GB. A Mac's GPU can use about 75% of its memory by default, and tools such as mflux load a model at 8-bit or 4-bit, which takes far less. So on a 16 GB Mac, start with FLUX.2 Klein 4B, Z-Image Turbo or Flux Schnell at 4-bit; Qwen Image's weights alone take about 11.5 GB at 4-bit, so give it 24 GB or more.

2What makes it fast

An image model refines a picture over a number of steps, and each step is a full pass through the model. A 4-step model like Flux Schnell or FLUX.2 Klein 4B makes a picture in a twelfth of the passes of Qwen Image at the 50 steps of its maker's example. Bigger pictures and higher precision take longer.

Unlike writing text, image generation is limited mostly by the GPU's compute, so chips with more GPU cores (Pro, Max, Ultra) are faster at it. Apple's M5 generation adds neural accelerators to each GPU core for this kind of work (Apple).

3Three ways to run them

With mflux, for example (mflux):

uv tool install --upgrade mflux
mflux-generate-z-image-turbo \
  --prompt "A puffin standing on a cliff" \
  --width 1280 --height 500 \
  --seed 42 --steps 9 -q 8

-q 8 loads the model at 8-bit, which roughly halves its memory against full precision. The first run downloads the model.

mflux, Draw Things and ComfyUI Desktop each have a listing in the AI App Store, next to other image and video apps.

4Licences and gated models

Open image models come with different licences. Flux Schnell and Z-Image Turbo are Apache 2.0; Flux Dev uses Black Forest Labs' non-commercial licence and is gated: Hugging Face asks you to accept it and sign in before it downloads (FLUX.1 dev, FLUX.1 schnell, Z-Image Turbo). Read the licence on the model card before using pictures commercially.

5Between pictures

A Mac that makes images in bursts is idle most of the day. MLX Node can use that time to run AI jobs for MLX Network within the limits you set and earn MLX Tokens; pause it when you want the GPU for yourself. Earnings aren't guaranteed: read the risks.

6Questions

Can a Mac with 16 GB generate images locally?

Yes, at a lower precision. FLUX.2 Klein 4B and Z-Image Turbo are small (4B and 6B parameters), and Flux Schnell's weights take about 6.7 GB at 4-bit. Tools such as mflux load models at 8-bit or 4-bit, which leaves room in the about 12 GB a 16 GB Mac's GPU can use.

What is the fastest image model on a Mac?

Among the models in the table, the 4-step ones: Flux Schnell and FLUX.2 Klein 4B, which is also the smallest. Speed also depends on picture size, precision and the number of GPU cores.

Does Stable Diffusion run on a Mac?

Yes: ComfyUI Desktop runs Stable Diffusion workflows on Apple Silicon Macs (M1 or later), along with newer families such as Flux. mflux focuses on the newer models listed above.

Can I use locally generated images commercially?

It depends on the model's licence, not on where it runs. Flux Schnell and Z-Image Turbo are Apache 2.0; Flux Dev's licence is non-commercial.