AI for UX · Labs · 5 min read
Generate images locally
A hands-on lab: generate images on your own computer with an open model. Three ways to start, how to read a model licence before you use the output, what we measured on a Mac, and what actually fixes hands and eyes (it isn't resolution).
Image models are no longer only in the cloud. Open models now produce images good enough for concept work, moodboards and placeholders on a recent laptop, with no per-image cost and no uploads. Two things matter more than the app you choose: the model's licence, which decides what you may do with the images, and how you write the prompt, which decides most of the quality.
What you'll be able to do after this lab:
- Pick a way to run an image model locally, with or without a terminal
- Read a model licence and decide whether you may use the output
- Generate an image and change one thing at a time to learn what matters
- Know the memory and time costs before you start
Products, versions and licences are as of November 2026. Check them again before you rely on them.
1. Three ways to start
A free app for Mac, iPad and iPhone that runs offline.
Easiest startCommands that run image models on Apple silicon (MIT).
Fast on a Mac; repeatableBuild workflows by connecting boxes (GPL-3.0).
Most control; steepest learning curve2. Read the licence first
- 1Which model?
Its exact name and version.
- 2Commercial use?
Allowed, limited by revenue, or never?
- 3Where?
Any excluded countries or regions?
- 4What else?
Attribution, no training on outputs, acceptance on download.
- 5Who approves?
Legal, for anything beyond personal tests.
The app's licence and the model's licence are different things. mflux is MIT, ComfyUI is GPL-3.0, but each model you download has its own terms:
- FLUX 3 Image Klein (Black Forest Labs, October 2026) uses the FLUX Kommunity License: free for non-commercial use, and commercial image generation is allowed only for "Qualifying Users" with under US$5 million gross annual revenue, counting affiliates. You must accept the licence on Hugging Face before downloading (model page).
- MiniMax-H3-Image lists the EU, the UK, South Korea and the USA as excluded territories in its licence (November 2026). If you're in one of them, don't use it.
- Some popular models are non-commercial only.
3. The terminal path on a Mac: mflux
mflux runs image models on Apple silicon with Apple's MLX library (MIT licence). In version 0.24, each model family has its own command.
# 1. Install (needs uv: https://docs.astral.sh/uv/)
uv tool install --upgrade mflux
# 2. Accept the model's licence on Hugging Face, then log in
hf auth login
# 3. Generate one image
mflux-generate-flux3 --model flux3-image-klein --quantize 8 --steps 16 \
--width 1024 --height 1024 --seed 7 \
--prompt "A small bakery shop window at dusk, hand-lettered sign reading 'Fresh bread'" \
--output bakery.png
In words: --quantize 8 compresses the model to use less memory, --steps 16 is how many refinement passes it makes, and --seed fixes the randomness so you can repeat the same image while you change one word of the prompt.
4. What we measured
On a MacBook Pro with an M3 Max chip and 36 GB of memory, with FLUX 3 Image Klein, 8-bit, 1024 × 1024 pixels and 16 steps:
| Measure | Result |
|---|---|
| Download | about 20 GB of model files |
| Time per image | about 85 to 93 seconds |
| Memory used by the whole process | about 14 GB |
| Text in the image (a shop sign) | spelled exactly |
What fixed hands and eyes. We compared the same prompt and seed with different settings:
| We tried | Result |
|---|---|
| Doubling the resolution | Twice the time, no better hands |
| Full precision instead of 8-bit | The same image, 6 GB more memory |
| 50 steps instead of 16 | Over ten minutes per image, no gain |
| Rewriting the prompt: figure large in the frame, "five fingers", "feet planted on the ground" | Fixed |
The lesson: framing beats settings. Describe what a good result looks like, the way you'd brief an illustrator. See AI for visual design: images and video and Prompt patterns for designers.
5. Common mistakes
- Using images from a model whose licence doesn't allow your use.
- Changing several things at once, so you can't tell what helped.
- Fighting quality with resolution and steps instead of a better prompt.
- Copying an artist's name into the prompt to imitate their style. It raises ethical and legal questions; see Ethics, bias and inclusion in AI design.
- Skipping the human review: check every image for distorted details and unintended content before it goes anywhere.
6. Vocabulary
| Term | Meaning |
|---|---|
| Diffusion model | A model that turns noise into an image, step by step |
| Steps | How many refinement passes the model makes |
| Seed | The number that fixes the randomness, so you can repeat an image |
| Quantised (8-bit) | Compressed to use less memory, with little visible loss |
| Licence | The terms that say what you may do with the model and its output |
7. Try it
No terminal? Use Draw Things (free, offline, Mac and iPad).
- Pick a model and read its licence with the questions in section 2.
- Generate one image with a fixed seed.
- Change one thing at a time: a word in the prompt, the framing, the size. Keep a note of what changed.
- Stop after five images and write down the one change that helped most.
Reflection questions: Which of your image needs (moodboards, placeholders, final assets) does a local model actually cover? Whose approval would you need before using a generated image in a real product?
Checklist
Learn more
- Tools: mflux · Draw Things · ComfyUI
- Models and licences: FLUX 3 Image Klein on Hugging Face
Books, courses and videos to go deeper
- How Diffusion Models WorkDeepLearning.AI with Sharon ZhouCourse · short online course (external site)
- Hands-On Generative AI with Transformers and Diffusion ModelsOmar Sanseviero, Pedro Cuenca, Apolinário Passos and Jonathan WhitakerBook · O'Reilly, 2024 (Amazon)
Book links go to Amazon. As an Amazon Associate I earn from qualifying purchases, at no extra cost to you.
Related
AI for visual design: images and video · Copyright and AI-generated assets · Run a model on your laptop · Prompt patterns for designers · Ethics, bias and inclusion in AI design