Image models
FLUX 3 Image Klein: API, References and Editing Checks
FLUX 3 Image Klein is a released BFL image model. This guide separates its hosted API from local weights and shows how to evaluate an edit before adopting it.
Black Forest Labs released FLUX 3 Image Klein on October 15, 2026. Its announcement describes a 5.6-billion-parameter image generator for text-to-image, editing and generation with up to ten reference images, including native 4K output. It is available through BFL and as open weights.
SJolt’s catalog checked October 19 does not list Image Klein. SJolt has separate FLUX.2 Pro image routes and FLUX 3 video routes. The shared family name does not make the following BFL request an SJolt request or turn the video model into an image endpoint.
Choose the hosted API or a local evaluation
BFL reports internal quality and latency comparisons in the announcement. They are publisher evaluations, not measurements reproduced here. Choose your own acceptance criteria before using those charts to justify a production change: exact lettering, product geometry, background stability and time to an accepted file can matter more than a single overall score.
The model card distinguishes a base checkpoint for adaptation from a distilled checkpoint used by the API. It also explains that the image generator, text encoder, autoencoder and optional prompt-expansion model contribute to a local setup. A quantized weight file’s size is not the complete memory requirement.
Use the hosted route if your immediate question is whether an output fits the brief. Use local evaluation when deployment control or model adaptation is part of the requirement. Record checkpoint revision, inference implementation, precision and prompt-expansion settings when comparing local results. This guide does not install weights or benchmark a GPU.
A small request to the official BFL API
The current BFL OpenAPI schema exposes POST /v1/flux-3-image-klein with x-key authentication. It requires prompt; the example explicitly chooses square 1k output and PNG. Running it submits a billable request to BFL. Set your own key privately on the server; no request was submitted for this article.
curl --fail-with-body 'https://api.bfl.ai/v1/flux-3-image-klein' \
-H "x-key: $BFL_API_KEY" \
-H 'Content-Type: application/json' \
--data '{
"prompt": "Editorial product photograph of an unbranded amber glass bottle on a pale limestone block. A single soft shadow falls right. Warm white background, front view, ample empty space above the bottle, no lettering.",
"aspect_ratio": "1:1",
"resolution": "1k",
"output_format": "png",
"seed": 19
}'The asynchronous response contains id and polling_url. Retain both and follow the returned polling URL according to BFL’s result contract. Do not resubmit merely because the image was not returned synchronously. A network interruption after submission requires checking the existing task before paying for another one.
For references, the BFL field is images, accepting a string or list. The first reference is the edit source and supplies the default size. Explicit width and height override resolution and aspect_ratio. These names differ from SJolt image-edit fields, so copy a complete provider contract rather than mixing parameters from tutorials.
Test preservation separately from a plausible image
The card advises full scene captions or prompt upsampling. Its reference CLI performs expansion by default, while the documented Diffusers pipeline needs that step supplied separately. References anchor content but do not guarantee unchanged structure. The card also warns that bounding boxes guide placement rather than acting as hard pixel masks.
For a bottle-background change, make the first reference the original product frame. Add a second reference only if it conveys a clear requirement such as surface texture. Describe the whole intended shot and name the specific change. Keep your source and requested edit together so the reviewer knows what must remain stable.
| Check | Compare against | Reject when |
|---|---|---|
| Identity | Original silhouette, cap and label | The product acquires a new shape or invented details. |
| Unchanged regions | Source at matching scale | A required region changes despite a local edit request. |
| Lettering | Approved text, character by character | Text is misspelled or replaced by plausible-looking marks. |
| Composition | Delivery crop and safe area | The final crop removes the subject or reserved copy space. |
If exact background preservation is essential, verify a documented preservation workflow and compare decoded pixels in the intended region. Do not infer a hard-mask guarantee from the word “box.” Keep final legal copy and editable typography in a layout tool when exact characters are required. A visually attractive candidate can still fail the production brief.
Open weights still have specific terms
The FLUX Kommunity License is not an unrestricted MIT or Apache grant. It permits non-commercial uses and provides a commercial-output exception for Qualifying Users, defined by combined gross annualized revenue below US$5 million. That exception has conditions, including output review or filtering and applicable AI disclosures. Other commercial uses require checking BFL’s licensing terms.
Before choosing self-hosting, identify who will operate the model, how outputs will be distributed and whether your use fits the actual grant. Keep the license revision with the selected checkpoint. Do not reuse an older FLUX family member’s license assumption or assume that downloading a file grants every intended commercial right.
Use the documented SJolt route when that is your requirement
For an image workflow already built on SJolt, FLUX.2 Pro is a separately supported option. Its text-to-image and image-edit operations have distinct routes; the latter accepts one to six image_urls. FLUX 3 on SJolt is video generation. Neither route is an alias for the BFL Image Klein endpoint described above.
Compare alternatives using one approved brief and a fixed review checklist, then record the exact model and settings for each candidate. Retain the chosen asset in your project and use a new request only for an intentional revision. This guide verifies source and contract details; it claims no generation result, quality winner or live integration of Image Klein.
Sources & further reading
Take the next idea into production.
Explore the models, test a workflow in the playground, and use the same request in your application.