FLUX.2 Klein 4B: generating and editing from one model on a consumer card, and where to run it
Apache 2.0 from Black Forest Labs, weights free, commercial use fine, nothing to sign. The number that matters for self-hosting is 13GB of video memory, which puts this on a consumer graphics card rather than data-centre hardware, using the Hugging Face diffusers library or ComfyUI. In your browser it runs on cnaps studio, included in the basic plan.
What makes it worth a post of its own is that one model covers three jobs. It generates a picture from a text prompt, edits an existing picture from a written instruction, and composes across several reference images at once. Those are usually three separate models with three sets of quirks, and here they share one set of weights and one interface.
The cost is that it is small. At 4 billion parameters it is a third the size of its faster sibling in this catalog, and nobody has published a number showing what that costs in quality. Seven image generation models sit in the cnaps studio catalog, and this is the lightest of the four that both make and change pictures.
What it actually is
A 4-billion-parameter rectified flow transformer. Rectified flow means the model learns a straighter route from noise to finished picture than older diffusion designs take, which is why a handful of steps is enough where those needed dozens.
The three modes are worth separating, because they need different inputs. Text to image needs only a prompt. Editing needs one reference image plus an instruction describing the change. Multi-reference composition takes several images plus an instruction, and is the mode for blending elements from different sources into one picture.
Settings are the same three as its sibling and no more. Diffusion steps run 1 to 8, defaulting to 4, and that is the only quality dial. A seed makes a result repeatable. Output is a square at 512, 768 or 1024 pixels. Worth knowing separately: the model ships with content filters built in, and outputs generated through the vendor's own hosted service carry provenance watermarking.
Try FLUX.2 Klein 4B in cnaps studio →
What is published, and what is not
Its model card has no benchmark table. No quality score, no evaluation against other generators, no measurement of what the smaller size costs. The vendor claims quality competitive with much larger models, and that claim has no number attached.
What is published is engineering: 4 billion parameters, the rectified flow design, under a second per image end to end, roughly 13GB of video memory, four steps, and Apache 2.0. Those are checkable facts about cost and capability rather than assertions about beauty.
The absence is common in this category. Of the seven generators here, only SANA-Sprint 1.6B, GLM-Image and DeepGen-1.0 publish scores, each on a different test, so none of them ranks this one. What you can act on is the specification: this is the model that runs on hardware you might own and does three jobs, and whether its pictures are good enough is a question your own prompts answer in an afternoon.
How it compares to the other image generators here
None of these four publishes a quality score on its own card, so the table compares what each one accepts and costs, and leaves the pictures to your own prompts.
The split is about inputs and hardware. Reach for FLUX.2 Klein 4B when the job involves an existing picture, when you want to blend references, or when the model has to fit on a card you own. Reach for FLUX Schnell when the input is only ever text and you would rather have the larger model. Reach for Z-Image-Turbo, SANA-Sprint 1.6B, DeepGen-1.0, GLM-Image or OmniGen2 when their particular strengths fit, and decide by running one prompt through several.
What to chain it with
It takes text, or an image plus text, or several images plus text, and returns one picture at 1024 pixels or less. Being able to accept images is what changes its position in a chain: it can sit in the middle rather than only at the start.
Two chains use that. For product work, BiRefNet (Subject) cuts your product out of its original photo and this model composes it into a new scene from a written description, which is a two-step route to a shot that never existed. For refinement, generate a base here, then send it to QWEN-Inpaint when one region needs replacing with a hard guarantee that nothing else moves. Either way, PiSA-SR or SMFANet+ at the end takes 1024 pixels up to print size.
Open FLUX.2 Klein 4B in cnaps studio, generate a picture, then feed that picture back in with an instruction to change one thing, and see how much of the original survives.
