How to Run FLUX 3 Image Klein on a Mac: Two Minutes a Picture, and Why That Changes How You Prompt

When Black Forest Labs (BFL) released FLUX 3 Image Klein on Thursday, the setup notes talked about NVIDIA graphics cards. Nothing about Macs. For the many illustrators, photographers and designers who work on a MacBook, that usually means a few weeks of waiting.

This time it took about a day. A community port called flux3-mlx appeared on October 16, and it runs Klein's full feature set on Apple Silicon: making images from text, editing your own photos, combining up to ten reference pictures, and placing things with boxes you draw.

There is one sentence in its README that every Mac owner should read before spending money on a bigger machine: "quantization buys memory, not speed." In plain words, the smaller versions of the model let it fit on a Mac with less memory, but they do not make it any faster. On the author's own test Mac, a picture takes about two minutes whichever version you pick.

Two minutes a picture is slow if you work the way most people use image tools, by rerolling until something good appears. It is perfectly workable if you change how you prompt. That shift is what this guide is about.

What you need

• An Apple Silicon Mac (any M-series chip). The tool is built on MLX, Apple's own framework for running AI models on these chips.
• At least 16 GB of memory. The port comes in two shrunken versions, called 6-bit and 8-bit after how much detail each stores per number. The README points 16 GB Macs to the 6-bit version and Macs with 24 GB or more to the 8-bit version, which stays closest to the original.
• Python 3.10 or newer and a few minutes in the Terminal.
• About 14 GB of free disk for the 6-bit setup. The picture model is an 8.4 GB download, and the prompt-writing helper (more on it below) is another 5.6 GB in its smaller form.

The model files download automatically from Hugging Face, the main public library for AI models, the first time you run it, with no Hugging Face account needed. Downloading them means you accept Black Forest Labs' FLUX Kommunity License, which is covered in the limitations below.

The steps

1. Install it. In Terminal:

pip install "flux3-mlx[pup]"

The [pup] part adds the prompt-writing helper. More on why you want it in a moment.

2. Make your first picture.

flux3-mlx "a ceramicist's studio at golden hour, shelves of glazed bowls" --upsample --out studio.png

On a 16 GB Mac, use the 6-bit version; the model card's example adds --model 6bit to the same command. If you are tight on memory, the README lists a --low-memory option that brings the 6-bit version's peak down to 9.6 GB.

3. Edit a photo you already have.

flux3-mlx "put a red bandana around the dog's neck" --image dog.png --upsample --out edit.png

Your photo sets the size of the result. To combine pictures, add more --image lines: "the woman from the first image wearing the jacket from the second" is the README's example.

4. Place things with boxes. Write a small file listing where each thing should go, as four numbers between 0 and 1 (top, left, bottom, right), and pass it with --boxes boxes.json. The README's example is hot-air balloons placed over a valley.

5. Keep the record file. Every image gets a matching .json file beside it with the final prompt, the seed, the size, the number of steps, the time taken and the memory used. Do not delete these. They are your notes.

How it works, in plain terms

Klein is really two models working in turn. The first is a prompt writer. It takes your short request ("a red bicycle in the rain") and expands it into the long, detailed scene description Klein was trained on. The second is the picture model, which turns that description into an image in 16 passes, called steps, each one cleaning up the picture a little more.

The --upsample flag runs the writer first. On the author's Mac it adds 8 to 10 seconds. You can skip it and write the long description yourself, which is a real option if you are good at describing images.

The port was checked against Black Forest Labs' own code. The author compared the outputs at each stage and reports that the full-precision version is "visually indistinguishable" from the company's published sample image. One practical result concerns the seed, the number that fixes the random starting noise every picture grows from. The same seed gives the same starting noise as BFL's own reference code, so the README says reference outputs can be reproduced on a Mac. If a collaborator running BFL's code finds a look you love, their seed and prompt give you a real chance of matching it.

The numbers, and what they mean for your day

All timings below are the author's, measured on an M4 Pro with 48 GB of memory, making 1024-pixel square images:

• A new picture: about 114 seconds with the 6-bit version, 115 with the 8-bit.
• An edit with one reference photo: about 200 to 270 seconds.
• Two references: about 290 seconds. Three: about 540 seconds, nine minutes.

No timings are published for 16 GB Macs or older chips, so expect yours to be the same or slower.

So, an hour gives you roughly 30 new pictures, or a dozen edits. Compare that with a fast cloud service, where you might make 30 in five minutes. The Mac route is free per image and private, because nothing leaves your machine. But it punishes the slot-machine habit of typing three words and pressing go twenty times.

Change how you prompt

Here is the workflow I would use with two-minute pictures.

Write the long version first. Short prompts give generic results with Klein, and that is a known trait, not a Mac problem. Instead of relying on the prompt writer every time, run it once on your idea, open the .json record, and read what it wrote. Then edit that description by hand: change the light, the lens feel, the materials, the mood. You are using the helper as a first draft, not as the author.

Change one thing at a time. Keep the seed from the record and change one phrase per run. With two minutes a picture, you learn more from four careful changes than from twenty random ones.

Batch the waiting. Write a short list of prompts in the evening, run them one after another, and look at the results in the morning. The record files will tell you exactly which prompt and seed made each one.

Save edits for when they matter. Edits with several reference pictures are the slowest job on the list. Draft your composition with plain text-to-image, then switch to an edit only for the final pass.

Do not go below 16 steps. It is tempting to cut steps to save time. The README says plainly not to: at 4 steps the output is "visibly under-denoised and grainy," meaning it looks unfinished and noisy, "in BFL's own code too."

Where it breaks

It is slow, and smaller files will not fix that. The author tested a 4-bit version and refused to ship it, because it "added 5–23% high-frequency grain and changed design details," meaning visible fine noise and altered details. That decision protects your images. It also means there is no faster version waiting.

Large images are impractical. The 6-bit model card calls 4K output "impractically slow" on current Macs.

It is unofficial. This is an independent community port, "not affiliated with or endorsed by Black Forest Labs." If something breaks, you are asking a volunteer, not a company.

The picture quality has the model's flaws. In the author's own 27-prompt test, reported on the 6-bit model card, some images had errors in hands and fingers, and some portraits came out with smooth, waxy skin. Boxes guide where things go but do not lock the rest of the picture in place, so edits can change areas you did not ask about.

The licence has conditions. The port's code is Apache 2.0, free for anything. The Klein weights are not. Under the FLUX Kommunity License, free commercial use of what you make is limited to businesses under US$5 million in annual revenue, other commercial use needs a licence from BFL, and the README notes it "comes with content-filtering and AI-disclosure duties." Read the licence before you deliver client work, especially since the model card says no safety filter is included.

It is two days old. Version 0.1.0 went up on October 16. Expect changes.

Where this leaves you

For years the honest answer to "can I run this on my Mac?" was "eventually, badly." Klein on Apple Silicon is neither eventual nor bad. It is slow, and it is faithful to the original.

That combination rewards a particular kind of creator: someone who knows what they want to see and can describe it. If that is you, try one evening of the long-prompt workflow above. Write five careful descriptions, run them overnight, and read the record files with your coffee. If the results look more like your work than anything you got from rerolling, the two minutes were never the problem.