FLUX.2 Klein Base 9B
The undistilled 9B base model of FLUX.2 [klein], for text-to-image and editing with up to 4 reference images, aimed at research and fine-tuning under a non-commercial license.
Overview
FLUX.2 [klein] 9B Base is the undistilled 9-billion-parameter checkpoint of the klein models from Black Forest Labs. Training skipped step and guidance distillation, and the model supports text-to-image generation as well as multi-reference editing. Black Forest Labs publishes it as open weights and lists creative exploration, fine-tuning, LoRA training and research among its intended uses.
Choosing a variant: pick this page for the largest undistilled klein model, for evaluation work or as the reference point for a fine-tune trained on the 9B weights. FLUX.2 Klein 9B is the same size with step distillation, defaults to 4 steps and guidance 1, and is meant for application use. FLUX.2 Klein Base 4B has under half the parameters and an Apache 2.0 license.
What you can build
- Landscape re-grades. One sample turns misty green hills into an autumn scene under a low sun, with fog still lying in the valleys.
- Golden-hour travel shots. From a hazy photo of moored fishing boats and the default sunset prompt, the model kept the boats in place and added a violet sky and a city skyline.
- Nature macro from text, such as a honeybee on lavender with pollen on its legs.
- Side-by-side tests against the distilled 9B: the same prompt and seed on both pages show what distillation changes at this size.
Inputs
The prompt is required and can run long: the default spells out sky colors, reflections, fog, boat textures and photographic style. Add up to 4 reference images, or none for text-to-image output.
It runs 30 steps at guidance scale 4 by default; steps accept 1 to 50 and guidance 1 to 20. Image size defaults to 1024 by 1024, with a ceiling of 2048 on either side. The samples field returns up to 8 images per request. The default seed of 0 draws a random seed on every run, so set a non-zero seed when you want to bring back an earlier image.
Outputs
In both edit samples the main subjects stay where they were in the source photo while light, color and atmosphere change. The hills edit also kept the layering of ridges and fog banks from the original.
Recommended settings
For parity with Black Forest Labs' example code for this checkpoint, set steps to 50 and keep guidance at 4. Expect more variety from one seed to the next than with the distilled 9B, so pick a non-zero seed before you explore and keep it once a composition works.
Limitations
Per the model card, do not rely on it for factual information, expect occasional errors in rendered text, watch for biases inherited from the training data, and plan for outputs that ignore parts of the prompt. Results depend heavily on phrasing. Custom LoRAs and fine-tuned checkpoints cannot be loaded on this page.
Safety & compliance
According to the card, the FLUX Non-Commercial License covering this checkpoint makes filtering or human review of NSFW and protected content mandatory for the 9B models. Black Forest Labs reports filtering NSFW content and known CSAM from pre-training data, plus safety fine-tuning and third-party evaluations. Prohibited uses include breaking the law, harming minors, misleading content, stalking or bullying, and intimate imagery made without consent. Only use a real person's photo as a reference if you have their consent.
Example prompts
Great starting points for FLUX.2-klein-base-9B.
API quick start
Run FLUX.2-klein-base-9B with a single API call.
{
"prompt": "A hyper-realistic wide shot of small fish…",
"inputImage": "https://your-cdn.com/input.png",
"steps": 30,
"scale": 4
}curl -X POST "https://api.wiro.ai/v1/Run/black-forest-labs/flux-2-klein-base-9b" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "A hyper-realistic wide shot of small fish…",
"inputImage": "https://your-cdn.com/input.png",
"steps": 30,
"scale": 4
}
JSONOther models you might like
Explore similar image generation models.



