FLUX 3 Klein vs FLUX 3 Max vs GPT Image 2.5: Test Results

What is FLUX 3 Klein?

FLUX 3 Klein is the smaller sibling of FLUX 3 Max, the flagship image model in Black Forest Labs' FLUX 3 family. It came out on October 15, 2026, and it's the first FLUX 3 image model with open weights, so you can download it and run it on your own GPU.

The interesting part is how little Black Forest Labs cut from the flagship. Klein keeps the same toolset: up to 10 reference images in one request, bounding boxes for placing elements, and native 4K output. On Black Forest Labs' internal evaluation, it ranks first among open-weights models in text-to-image, editing and multi-reference generation, ahead of larger open models. Compared with the previous FLUX.2 [klein] 9B, it scores 91 Elo points higher in text-to-image and 70 points higher in editing, with 38% fewer parameters.

The two FLUX 3 models side by side:

The gap grows with resolution: at 1K, Max costs about five times more than Klein, and at 4K, almost 17 times more.

How we tested

We put the models through six tests built around what Black Forest Labs and OpenAI show off in their own launch demos: an infographic with exact text, the same infographic translated into Portuguese, one character across three scenes, an edit that must leave everything else untouched, a magazine cover laid out with bounding boxes, and a room furnished from five reference images. We picked tasks where image models usually slip up, from small text to stray pixels.

All three models ran through their APIs with the same prompts and input images, at about 1 megapixel. We kept the first result of every run, with no rerolls, and timed every generation, since speed is the main thing Klein promises.

We'll also compare GPT Image 2.5 Flare against FLUX 3 Klein, arguably the model to beat among lightweight AI image generators when it comes to style, text rendering, visual fidelity, and instruction following.

One thing we found (spoiler alert) is that GPT Image 2.5 Flare sacrifices some accuracy for creative flair, making it less suitable for certain tasks. More on that later.

Test 1. Technical infographic with exact text

The first task is an infographic about how a mechanical watch keeps time, showcasing a cutaway of the movement with five labeled parts, five numbered steps with sentences written word for word, and arrows that follow the flow of energy.

FLUX 3 Klein's version is on the left, FLUX 3 Max's on the right and GPT Image 2.5's below:

Technically, FLUX 3 Max did the best job: it labeled all five parts of the movement, gave each step its own frame, and used arrows to connect the steps to the parts they describe. Of the three, it produced the most accurate visual explanation of the concept.

Visually, though, FLUX 3 Klein produced the cleanest and most appealing result. That's particularly surprising considering it's competing against GPT Image 2.5, a much more advanced model.

Other than that, all three models rendered the five sentences without a single mistake.

If you want to replicate this, here's the test prompt:

Translating the infographic into Portuguese

Next, each model translated its infographic into Portuguese without changing anything else. The translated versions are in the same order:

GPT Image 2.5 translated every line into readable Portuguese, but it paraphrased a few sentences. Step 3 gained "ao longo do mecanismo" ("along the mechanism"), and step 2 calls the mainspring "barrilete principal", the barrel that holds it.

FLUX 3 Max translated the title, the part labels and four steps cleanly. Step 3 turned into nonsense ("ospea de npr'ado pereongeno"), and steps 3, 4 and 5 moved to different frames than in the English version.

FLUX 3 Klein translated the title, but the small text broke apart: "a aperta a ·pirital", "ma pincipal", and the label "crown" stayed in English. The numbering broke too. Step 3 appears twice, and the escapement is described in two different steps.

So, Klein isn't really usable when you need to modify a large amount of text, particularly when languages other than English are involved.

The prompt:

Keeping one character consistent across three scenes

The third task follows one character through a short story, which demonstrates multi-turn consistency. Meet Pip — he's a small brass robot we'll be using for this test:

FLUX 3 Klein and FLUX 3 Max show impressive character consistency. Pip looks exactly the same across all frames, and so does the art style. Whether that's a good thing depends on what you're trying to achieve.

GPT Image 2.5, on the other hand, seems to have picked up on the brushstroke texture of the input image and amplified it, making the effect more pronounced with each frame. Technically, that's a failure in terms of what we were testing, but it's hard to deny how good the final image looks. Although, for some reason, Pip also grew an antenna that wasn't in the original image or the prompt.

If you want to run this test, here are the prompts in order:

Editing one object without modifying anything else

How often do you use an AI image generator to change something in an existing image? We run into this use case all the time, especially when creating before-and-after comparisons. Ideally, when you ask a model to edit one specific part of an image, everything else should stay exactly the same. But that's not always what happens. Some models end up changing far more pixels than the task requires. Black Forest Labs, however, claims that FLUX 3 preserves all pixels outside the edited area.

To test this, we used a photo of a home office desk and asked each model to remove the white mug and replace the white chair with a dark green leather one, leaving everything else untouched.

Below each result is a map showing which pixels changed. We counted a pixel as changed if at least one of its color channels differed from the original by 32 or more on a scale of 0 to 255. All changed pixels are highlighted in red:

FLUX 3 Max changed 10.2% of the pixels, almost all of them around the mug and chair. GPT Image 2.5 changed 12.3%, also mostly in those areas. FLUX 3 Klein, however, changed 17.2%, with small differences scattered across almost the entire image.

That's because the whole image shifted by a pixel or two, causing the edges to no longer line up with the original. This becomes a problem if you want to show a quick before-and-after comparison: instead of transitioning smoothly, the image will visibly jump between the two versions.

The exact prompt:

Laying out a magazine cover with bounding boxes

When a design needs every element in a fixed spot, like a magazine cover with room for the title, FLUX 3 lets you draw a box for each element and describe what goes in it. Is this more accurate than simply describing changes?

For this test, we used the boxes to outline where FLUX 3 Klein should make changes. The interface is a little different for FLUX 3 Max, so we added JSON at the end of the prompt with positions. GPT Image 2.5 has no such feature, so it got the same JSON plus one sentence explaining the format.

The dashed lines show the boxes we requested:

FLUX 3 Klein and FLUX 3 Max placed every element inside its box with pixel-perfect accuracy. Unfortunately, no magic happened for GPT Image 2.5. It still placed the elements in roughly the right areas, but they were consistently a little out of bounds.

Here's the prompt with the aforementioned JSON:

The boxes, as [top, left, bottom, right] on a 0 to 1000 grid:

Furnishing a room from five reference images

Lastly, the multiple reference test: the models get an image of an empty living room and 4 objects, and they need to place the objects in the correct perspective and scale without changing what they are. The five references are at the top, and the results are below:

The prompt:

Interestingly, FLUX 3 Klein zoomed in on the corner of the room where it arranged the furniture. That's an unexpected side effect, and it suggests that the model might need more explicit instructions to preserve the original camera angle and perspective when working from a reference image.

GPT Image 2.5, on the other hand, kept the original composition intact but also chose to arrange the furniture in the far corner of the room (the same corner Klein picked, actually).

That's an interesting detail. From an interior design perspective, placing furniture in the middle of an empty room wouldn't make much sense, so both models seem to have used their understanding of real-world spaces to create a more natural arrangement. The difference is that GPT Image 2.5 preserved the original composition at the expense of making the furniture harder to see, while Klein changed the framing to give the objects more prominence.

Speed and price

Speed is the main reason to pick FLUX 3 Klein, and the difference is easy to feel. Here's a table that shows how fast it generates images as compared with FLUX 3 Max and GPT Image 2.5 Flare:

In other words, on average, Klein returned an image in about 6 seconds, more than four times faster than GPT Image 2.5 Flare and six to seven times faster than FLUX 3 Max.

The type of task matters too. Klein made a new image from text in 4 seconds, simple edits took 6–7 seconds, and combining five reference images into the living room took 12. FLUX 3 Max varied the most, from 23 seconds for the first robot picture to 67 seconds for the Portuguese translation. GPT Image 2.5 stayed between 21 and 32 seconds no matter what we asked.

FAQ

Below are short answers to common questions about FLUX 3 Klein.

What is FLUX 3 Klein?

FLUX 3 Klein, full name FLUX 3 Image Klein, is a compact image model from Black Forest Labs, released on October 15, 2026. It has 5.6 billion parameters and open weights. It generates images from text, edits them, combines up to 10 reference images, places elements with bounding boxes and outputs native 4K. There are two versions: a distilled one that needs only 16 steps per image, and a Base version with 50 steps for fine-tuning and LoRA training.

What's the difference between FLUX 3 Klein and FLUX 3 Max?

FLUX 3 Max is the larger flagship model and is available through an API, while Klein's weights are free to download. Both support 10 references, bounding boxes and 4K output. In our tests, Klein was about six times faster and five times cheaper, and it matched Max on bounding box layouts and character consistency. Max was better at editing without touching the rest of the image, at keeping the room unchanged when combining references and at long text in infographics.

Is FLUX 3 Klein free for commercial use?

Only for smaller companies. Klein is released under the FLUX Kommunity License, which allows non-commercial use and commercial image generation for users with less than $5 million in gross annual revenue. Larger companies need a commercial license from Black Forest Labs. The license also doesn't automatically allow commercial fine-tuning or redistributing the model.

Can you run FLUX 3 Klein locally?

Yes. The weights are on Hugging Face, and there's a version for ComfyUI. The INT8 build runs on NVIDIA RTX cards with 8 GB of video memory: a 1 MP image takes about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070. Black Forest Labs recommends never running Klein without its prompt upsampler, a separate 9B model that needs about 28 GB of GPU memory. You can replace it with another language model by using the system prompts Black Forest Labs provides.

Is FLUX 3 Klein better than GPT Image 2.5?

In our six tests, Klein beat GPT Image 2.5 Flare at placing elements in bounding boxes and at keeping a character's look across scenes, and it generated images about four times faster. GPT Image 2.5 was better at translating the text in an image, changed fewer pixels outside the edit and kept the original room unchanged when combining references.