Updated
What is FLUX 3 Image Klein?
FLUX 3 Image Klein (the flux 3 klein API, written "FLUX 3 [image klein]" by Black Forest Labs) is a 5.6-billion-parameter image transformer released on October 15, 2026. It does three jobs in one model: text-to-image, image editing, and multi-reference generation from up to four input images. Output goes up to native 4 MP, and the default schedule is 16 steps, because this is the step-distilled build. Text goes through a Qwen3-VL-4B-Instruct encoder, which is a big part of why it reads prompts and reference images better than FLUX.2 Klein did.
The weights are open under the FLUX Kommunity License (commercial use allowed under five million dollars in revenue), in BF16, FP8, INT8 and NVFP4 builds. The INT8 build fits an 8 GB RTX card. The undistilled sibling, FLUX 3 Image Klein Base, runs 50 steps with CFG guidance and a negative prompt and is meant for fine-tuning and LoRA training. For API work, this page is the one to use.
On UnificAlly the route takes a prompt, seven aspect ratios (21:9 through 9:21), four resolution presets (0.6, 1, 2 and 4 MP), up to four reference images, steps from 1 to 50 and a seed. Billing is per finished image by resolution, with reference-image runs priced higher.
Key features of FLUX 3 Image Klein
Five things you will actually touch in a production run.
Generation, editing and references on one route
Send a prompt alone for text-to-image. Add one reference in image_urls for an edit, or up to four for multi-reference work. Same request shape, same model id, no mode switch.
Native 4 MP output
Four presets: 0.6, 1, 2 and 4 MP. The 4 MP preset renders at full resolution rather than upscaling a smaller image, so fine detail like engraved serial numbers and brushed metal holds up at 100% zoom.
Text that renders correctly
Two lines of signage in one prompt, both spelled right, with the paint wear and rust the prompt asked for. Good for mockups, posters and product shots that carry real words.
5.6B open weights, 16 steps by default
Step-distilled, so 16 steps is the tuned default and 1 MP returns in about 3 seconds on our route. Turn steps up to 50 when you want to trade speed for a cleaner result, or down for drafts.
Seven aspect ratios and a seed
21:9, 16:9, 4:3, 1:1, 3:4, 9:16 and 9:21 on one prompt, so one brief covers banner, feed and story crops. Pass a seed and the same inputs return the same image every time.
What FLUX 3 Image Klein is best for
Product and catalog editing
Pass the product shot as a reference and describe the change: new background, new season, new color. Layout and lighting survive the edit, so the result still matches the rest of the catalog.
Interactive design tools
About 3 seconds at 1 MP keeps an editor responsive. Draft at 16 steps, then re-render the picked frame at 4 MP for export.
Multi-reference composites
Up to four references in one call. Combine a product, a model and a scene without a compositing step in between.
Signage, posters and mockups
Text renders legibly across two lines at 1 MP. Useful for storefront mockups, packaging drafts and ad comps that carry real copy.
Cross-platform asset sets
Seven aspect ratios on one prompt plus a locked seed. Generate the banner, the square and the story crop from the same brief and keep them visually consistent.
Teams that also run locally
The open weights mean a local ComfyUI pipeline and the API call produce the same model output. Prototype on an 8 GB card, run production through the API.
FLUX 3 Image Klein use cases
Build a product-page editor that takes the existing hero shot as a reference and swaps the season, backdrop or colorway from a short prompt, then re-renders the approved frame at 4 MP for the CDN. Run a storefront mockup generator where the shop name and tagline come from a form field and land on the sign in the image, spelled correctly. Set up a social asset pipeline that fans one approved prompt and seed across 16:9, 1:1 and 9:16 so the campaign looks the same on every surface. For a design tool, draft at 1 MP and 16 steps for near-instant feedback, and only spend the 4 MP render on the frame the user exports.
What FLUX 3 Image Klein can do
House-generated examples on this route, all at seed 7 and 16 steps.
Layout-preserving edits from a reference
The FLUX.2 Klein 9B cover went in as the reference with a prompt asking for deep winter. The typographic layout, framing and camera stay exactly where they were; the letters are now pressed into snow under cold blue dusk light. This is the edit behavior you want for catalog and brand work.
Native 4 MP detail
A 2048x1504 render from the 4 MP preset. Blued screws, ruby jewels and the engraved serial numbers on the bridges are all rendered, not smeared, and raking light reads as real brushed steel. Returned in about 8 seconds.
Two lines of text, both spelled right
A weathered enamel sign with "FLUX 3 KLEIN" in bold block letters and "OPEN WEIGHTS" below it. Both lines come out correct at 1 MP, with the rust and paint wear the prompt described. Text rendering is a known weak spot for small image models; this one holds.
Skin, hair and natural light at 1 MP
An editorial portrait with freckles, silver-streaked curls and soft window light from the left. Skin texture and individual hair strands survive the 16-step distilled schedule, and the film grain is subtle rather than noisy. About 3 seconds on our route.
FLUX 3 Image Klein vs FLUX.2 Klein
Black Forest Labs' own evals (vendor-reported, so read them as a claim rather than an independent benchmark) put FLUX 3 Image Klein first among open-weight models in text-to-image, multi-reference generation and image editing. Against FLUX.2 Klein 9B it gains 91 Elo in text-to-image and 70 Elo in image editing with 38% fewer parameters. BFL also says it beats Ideogram V4.5 and FLUX.2 Pro on the same evals. Arena.ai had no FLUX 3 entry as of October 19, 2026, so there is no independent rank yet.
In practice the differences you will notice: FLUX.2 Klein had no steps control and took up to ten references; FLUX 3 Image Klein exposes steps 1 to 50 and takes four references, with better prompt and text adherence per image. If you are on FLUX.2 Klein 4B or 9B today, the request shape is nearly the same, so moving is a model-id change plus a reference-count check.
FLUX 3 Image Klein API examples
Call FLUX 3 Image Klein from any language by POSTing to /v1/tasks. Full parameter docs live at docs.unifically.com/models/image/flux/flux-3-image-klein.
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "black-forest-labs/flux-3-image-klein",
"input": {
"prompt": "A weathered green enamel sign on a brick wall that reads FLUX 3 KLEIN in bold white block letters, morning sun",
"aspect_ratio": "16:9",
"resolution": "1mp"
}
}'
Successful submission returns a task_id. Poll GET /v1/tasks/<task_id> or set a callback_url on the request to receive the finished image URL.
FAQs
People also ask about FLUX 3 Image Klein
FLUX 3 Image Klein is Black Forest Labs' 5.6-billion-parameter image model, released October 15, 2026 with open weights. One model handles text-to-image, image editing and multi-reference generation, renders natively up to 4 MP, and runs its default 16-step distilled schedule. On UnificAlly it is the black-forest-labs/flux-3-image-klein route.
A prompt, aspect_ratio (21:9, 16:9, 4:3, 1:1, 3:4, 9:16, 9:21), resolution (0.6mp, 1mp, 2mp, 4mp), up to four reference images in image_urls, steps from 1 to 50 (default 16), and an integer seed. There is no guidance, negative prompt or prompt upsampling on this model.
Pass the image you want to change as a reference in image_urls and describe the change in the prompt. The model keeps the composition, layout and lighting of the source and applies the edit on top. Up to four references can be combined in one call for multi-reference work, such as placing a product into a scene.
Klein is the step-distilled model, tuned for 16 steps and fast turnaround. Base is the undistilled model behind it, running 50 steps with CFG guidance and a negative prompt. Base exists for fine-tuning and LoRA training; Klein is the one to call in production.
Yes. Black Forest Labs released the weights under the FLUX Kommunity License, which allows commercial use for companies under five million dollars in revenue. Builds come in BF16 (11.2 GB), FP8 (5.6 GB), INT8 and NVFP4; the INT8 build runs on an 8 GB RTX card. The UnificAlly API serves it with no download or GPU.
On the UnificAlly route a 1 MP text-to-image call returns in about 3 seconds. Runs with a reference image take about 5 seconds, and a 4 MP render about 8 seconds. Black Forest Labs quotes 4.54 seconds for text-to-image and 7.18 seconds for an edit at 1 MP on its own API.
Per finished image, by resolution preset. Runs that include reference images are billed at a higher rate than plain text-to-image. Failed generations are not charged. The current rates are on the pricing page.
Yes. Send an integer seed with the request. The same prompt, references, aspect ratio, resolution, steps and seed return the same image, which is what you want for regression tests and for locking a look across a batch.
