FLUX 3 Image Klein Runs on an 8 GB Graphics Card. Its Prompt Writer Is the Part That Does Not Fit.
The headline on Black Forest Labs' new open image model is a number most creators can act on: 8 GB. FLUX 3 Image Klein, released on October 15, makes images from text, edits photos you already have, and combines up to ten reference pictures, and the company says a compressed version "runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM." That is a mid-range gaming card from a few years ago.
Then you open the setup guide and find a sentence in bold: "Do not use the model without a prompt upsampler."
The prompt upsampler is a second model. Its job is to turn your short request into the long, detailed scene description Klein actually wants. Klein itself has 5.6 billion parameters, a rough measure of size. The writer has 9 billion. So the helper is bigger than the painter, and the 8 GB figure describes the painter.
That is the part launch coverage skipped, and it changes how you should set this up. I think it also makes Klein more interesting, not less, because it hands you a choice most image tools hide: who writes the description of your picture.
Why a free open model matters this week
Klein arrived on the same day as Krea 3, a paid editing model from Krea, and the two pitch nearly the same thing. Draw a box around one part of a picture, describe the change, and keep everything else. Krea charges six cents per image at 2K. Klein is free to run on your own card if you or your company earn under US$5 million a year, under its FLUX Kommunity License. Above that you need a licence or BFL's paid service, which BFL says starts at "$0.009 per 1K image."
For a product photographer, a social designer or an illustrator who already owns a decent graphics card, that is the whole pitch. Edit one region of a client image, as many times as you like, without a meter running. BFL's own timings, using the fast version on a compressed file, are about 4 seconds per image on an RTX 5090 and under 9 seconds on an RTX 4070. Those are the maker's numbers, so treat them as a ceiling until you see your own.
The open release also means other people build on it immediately. By the morning after launch, a search on Hugging Face for the model name returned nearly forty community projects: smaller files for smaller cards, versions for Apple Silicon Macs, an add-on for extending photos past their edges, and, tellingly, a shrunk copy of the prompt writer. The community went straight for the part that does not fit.
The two-model design, in plain terms
Here is what actually happens when you type "a red balloon over a wheat field" into Klein.
First, the writer reads your request (and any reference photos, if you are editing) and expands it into a full caption: the subject, the setting, the light, the materials, the camera feel, where things sit in the frame. Then the picture model reads that caption and paints. The writer runs once per image, before any painting starts.
BFL is direct about why it built it this way. The model card says Klein "takes full scene captions, for text-to-image, editing and layout with boxes alike," and that "a short prompt sent as typed gives generic output." In other words, Klein was trained on long descriptions. Hand it a short one and it fills the gaps with the most average possible answer.
This is why the 8 GB story needs a footnote. BFL's ComfyUI package includes a compressed writer file next to the compressed picture file. The card warns that a compressed file's size "is not a measure of total GPU memory use," and the community repo of smaller Klein files calls the 9B writer "demanding on low-VRAM systems" and suggests skipping it while you test. On a big card both sit in memory together. On an 8 GB card, they will not both fit comfortably, and something has to give.
You have three honest options.
Run the writer somewhere else. The writer is a language model, the same kind of thing that powers a chatbot, so it runs in the free desktop apps people already use for chatbots. A community conversion on Hugging Face packages it for LM Studio, Ollama and llama.cpp, with a recommended file of 5.63 GB. It can run on your computer's main processor and memory, or on a second machine, while the graphics card stays free for pictures. BFL's own setup guide recommends the same split for smaller cards, when the two do not fit together: run the writer first, then pass its caption to Klein with the writer switched off.
Skip the writer and write the caption yourself. BFL's README says not to switch the writer off "unless the prompt already is an upsampler caption," and it calls captions made without one "noticeably weaker." Read that as: if you write like the writer, you do not need the writer.
Use someone else's card. The free demo on Hugging Face runs the whole thing, writer included, on shared hardware.
Why writing your own caption is the creative move
The second option is the one I would push working creatives toward, at least some of the time.
When a writer model expands "a red balloon over a wheat field," it makes dozens of decisions for you. Time of day. Lens. Whether the wheat is green or gold. Whether there is a horizon. Those are exactly the decisions a photographer or art director gets paid to make. An automatic writer makes them the same way for everyone, which is how AI images end up looking alike.
Writing the long caption yourself is slower, and it is also where your taste goes into the picture. It is closer to writing a shot description for a crew than to typing a search query. If you already write creative briefs, you have most of this skill.
The workable middle ground: let the writer draft the caption, read it, and rewrite the parts that matter to you before Klein ever sees it. Running the writer separately makes that easy, because its output lands as text you can edit instead of vanishing inside a pipeline.
Put this into practice
1. See the difference before you install anything. Open the free Klein demo on Hugging Face. Upload one of your own product or portrait photos and ask for a small change in one short sentence. Then ask for the same change with a full paragraph describing the scene as you want it. Compare. The demo says it can improve your prompt automatically, so pay attention to how much your own paragraph changes the result compared with your one-liner. It runs on shared free hardware, so expect queues.
2. Install the ComfyUI version. ComfyUI is the free node-based app most local image work runs in. BFL published official templates on launch day, including one for editing, and a ComfyUI file repo with the picture model, the text reader and the writer in compressed form. You have to log in to Hugging Face and accept the licence before the files download. The files marked "sd" appear to be the fast 16-step version (the repo does not spell out the label), and that is the one to start with.
3. Move the writer off the graphics card if you have 8 to 12 GB. Install LM Studio or Ollama and load the community writer file at its suggested size. Its card notes separate modes for new images, edits and box layouts, each chosen by an instruction you give it, so copy those from the card rather than guessing. Paste its caption into ComfyUI and bypass the writer node there.
4. Build a caption template you reuse. A simple one: subject and what it is made of; where it sits in the frame; background; light source and time of day; camera distance and lens feel; mood; any text that must appear, in quotes. Fill it in for your next three jobs. You will notice which lines change the picture most, and those become your house style.
5. Edit with a box, then check the edges. For regional edits, Klein accepts boxes that say where something goes and how big it is. Use one, then lay the result over the original in your editor with a Difference blend. Anything that is not black changed.
Where it breaks
The 8 GB claim is about the picture model. Everything above follows from that. If you run the full package as shipped, the writer comes along, and your card has to hold both. Budget for the writer, not just the painter.
The writer files most people will use are unofficial. The small writer for LM Studio and Ollama is a community conversion, not BFL's file. A shrunk writer may write weaker captions. The Mac versions that appeared overnight are community builds too.
Boxes guide, they do not lock. The model card says it plainly: "Boxes guide placement and scale; they do not constrain pixels to a hard mask," and edits can change content outside the requested region. The Difference check in step 5 is not optional on client work.
Every quality claim is BFL's own. The launch post ranks Klein first among open models and ahead of some paid ones, based on BFL's internal evaluations. Your own images are the only test that counts.
The licence has a ceiling. Under US$5 million a year in revenue, commercial use is free. Every community add-on and compressed file inherits the same terms. If you work for a large studio or agency, check before you deliver anything made with it.
It is a day old. Files, templates and memory advice will change over the next few weeks, and some community uploads will not work with the official tools. Expect friction.
Where this leaves you
The pitch for Klein is free, private, box-guided editing on a card you already own, and that pitch mostly holds. What it leaves out is that the model was built to be fed by another model, and that second model is the heavier one.
You can treat that as a hardware problem and solve it by moving the writer off the card. Or you can treat it as an opening. A model that needs a long, specific description of every picture is a model that rewards people who know how to describe pictures. Try one week of writing your own captions for real jobs and keep the three you like best. If the results look less like everyone else's, the 9B writer was never the part you needed.
Medium metadata
Title: FLUX 3 Image Klein Runs on an 8 GB Graphics Card. Its Prompt Writer Is the Part That Does Not Fit.
Subtitle: Black Forest Labs' new open image editor fits on a mid-range card, but it ships with a larger helper model that writes your prompts, and where you run that helper (or whether you replace it with your own captions) decides how well it works for you.
Tags: Artificial Intelligence, AI Art, Photography, Image Editing, Open Source
Estimated read time: 8 minutes