FLUX 3 Klein Is Here: What It Does Better Than FLUX.2

Black Forest Labs, the lab behind FLUX, released FLUX 3 Klein today. It's the small, fast model in the new FLUX 3 family (“klein” is German for small), and it follows FLUX.2 Klein, the model behind BestPhoto's quick image generation and photo edits.

We gave both the same requests. The new one, Klein 3, wins most of them, by a lot when you're editing a photo of a person. The old one, Klein 2, still wins a couple, and we'll show you which.

Edits finally leave the rest of your photo alone

You know the feeling. You ask an AI for one small change, and the whole picture quietly shifts: the furniture moves, or the person comes back looking like their cousin.

We asked both models to turn a sunny living room into a cozy rainy evening: lamp on, rain on the window, a sleeping orange cat, books and tea on the table, and “keep the furniture and layout the same.” Both got the lamp, the cat, the books and the tea. Klein 2 also made it rain inside the living room, with streaks over the sofa and a wet table. Klein 3 put the rain on the glass and left the room alone.

That's the quieter upgrade. Klein 3 reads a request the way a person would: rain on the window means rain on the glass.

Faces matter most. We asked for four changes to one portrait at once: a short curly bob, small gold hoops, a denim jacket and a sunny café instead of the gray wall, plus “keep her face exactly the same.”

Both made all four changes. Klein 3 gave back the same woman. Klein 2's came back younger, with a different nose and mouth, more like her little sister.

Klein 2's curls are curlier and its café is sunnier. But if you wanted to see how a bob looks on her, only one of them answers the question.

Cleaner signs, and the same face from three photos

Words are the next clear win. Ask for a yard-sale poster with three lines of writing, and Klein 3 spells every one. Klein 2 gets most of it, but a single typo like “Maple Strceet” ruins a sign.

Mixing photos is another. Hand it a photo of a woman, a jacket and a street, and ask for her in that jacket on that street. Only Klein 3 brings back the same woman.

Plain photos can feel more like a real moment, too, like the rainy street at the top of this post.

SpellingKlein 3 winsThree photos inKlein 3 winsNight streetKlein 3 winsWoodblock printKlein 2 wins

Prompt“A poster for a yard sale that says "BIG YARD SALE", "Saturday 8am - 2pm" and "42 Maple Street", with hand-drawn pictures of a lamp, a bike and a teddy bear”

Klein 2

Klein 3

The last line, up close

Klein 2“Strceet”

Klein 3“Street”

Klein 2's poster has more homemade charm, but only Klein 3's sends people to the right street.

It puts things where you tell it

Klein 3 can also follow a layout. Mark a box for each thing, say what goes in it, and the model puts it there: the latte top center, the croissant bottom left, the phone on the right, the sunglasses in the corner.

Describe the same layout in words alone and you get close, but things drift. The latte grows too big and the sunglasses slide half out of their spot.

You don't need boxes to use this. Plain position words in your request (“top center,” “bottom left corner”) get you most of the way, which helps when you need room for a headline or want a product in one corner.

Faster on big pictures, not everyday ones

Small and fast doesn't mean faster than before, so don't expect everyday pictures on BestPhoto to arrive twice as quick tomorrow. At normal sizes the two models finish about together, with Klein 2 a hair quicker.

The bigger the picture, the further Klein 3 pulls ahead. At four times the everyday size, it finishes in about half the time Klein 2 needs. And it goes all the way to 4K, in about 14 seconds.

Where it still slips

Strict art styles are one. In the woodblock print above, Klein 2's fox looks like the real thing. For styles like that, Black Forest Labs suggests describing the look instead of only naming it: the colors, the line work, the texture.

Busy scenes can blur together. In a reading nook packed with plants, books, a sleeping cat, a teapot, fairy lights and a record player, Klein 3 merged things: the teapot stands on a vinyl record, and one “record player” looks more like a pan. Klein 2 kept every object separate.

And both still get carried away. Asked to put Santa hats on the two kids in a family photo, both put hats on Mom and Dad too. Count the hats before you print the card.

How to ask so you get what you meant

With this model, the way you write a request changes the result a lot, Black Forest Labs says. Three habits help.

Say what should stay. A bare “make it blue” can reframe the whole shot. End an edit with what to keep: “keep her face exactly the same,” “keep the furniture and layout the same.” Both edits above were written that way.

Describe the whole scene. Short prompts give generic pictures. Instead of “dog at the beach,” try “a golden retriever leaping for a frisbee at the water's edge, late afternoon sun, shot from low down.”

Put words in quotes. Write the text you want on a sign or poster in quotation marks, spelled exactly. Text you didn't ask for, like the fine print on a label, can still come out as gibberish.

Quick questions

The results in this post are real outputs from FLUX.2 Klein and FLUX 3 Klein, the first try for each request. The people in the input photos were made with AI and aren't real.