ImageKleinStart creatingOne model for generating and editing
FLUX 3 Image Klein takes a prompt plus up to 10 reference images and returns one image. The same model does all three jobs:
- Text-to-image — a new image from a description, including legible text inside the image.
- Image editing — change one image with an instruction. Through the API, “pixel restoration” keeps every pixel you didn't ask to change exactly as it was.
- Multi-reference generation — combine people, products, styles or objects from up to 10 images into a new one.
It inherits three features from the larger FLUX 3 Image model: native 4K output, pixel-precise editing, and bounding boxes that tell the model where an element should go or confine an edit to one region (how bounding boxes work).
The checkpoints
| Checkpoint | Steps | Guidance | Use it for |
|---|---|---|---|
| FLUX 3 Image Klein (step-distilled) | 16 | none | Fast generation and editing, self-hosting; the checkpoint behind the API |
| FLUX 3 Image Klein Base | 50 | 3.5 | Fine-tuning, LoRA training, research |
| FP8 / NVFP4 / INT8 builds | same as above | Less memory; INT8 is the 8 GB GeForce option | |
| Prompt upsampler (flux-3-image-klein-pup) | — | Rewrites short prompts into full captions (≈18.8 GB download) | |
The image generator is 5.6B parameters: 11.2 GB in BF16 and 5.6 GB in FP8. Prompts are read by a separate text encoder, Qwen3-VL-4B-Instruct (Apache-2.0).
How good is it?
Black Forest Labs' own blind evaluations put FLUX 3 Image Klein first among open-weights models in text-to-image, multi-reference generation and image editing — ahead of larger open models such as Qwen-Image 2.1 (7B) and Krea 2 (12.9B), and above the closed Ideogram V4.5 and FLUX.2 [pro]. Against FLUX.2 [klein] 9B it gains +91 Elo in text-to-image and +70 Elo in editing.
These are BFL's internal benchmarks. VentureBeat notes the charts don't include GPT Image 2.5 or Nano Banana 2.1 and don't publish prompts or sample sizes, so treat them as the maker's claims until independent tests appear.
Why it matters
- Local and private: it runs on a consumer GPU or a laptop, so images never leave your machine — useful for data-sovereignty rules and on-device editing.
- Cheap to adapt: the Base checkpoint is built for LoRA fine-tuning (Ostris AI Toolkit, Diffusers, NVIDIA NeMo AutoModel); serving a LoRA adds about 5% to generation time.
- Cheap to call: through the API it starts at $0.009 per 1K image, versus $0.048 for the full FLUX 3 Image.
Safety
BFL filtered the pre-training data, fine-tuned the model to suppress harmful outputs, red-teamed both checkpoints internally and had the external partner Cinder evaluate them before release. The license forbids, among other things, non-consensual intimate images, CSAM, biometric processing, surveillance and violating people's publicity or digital-replica rights.
Where it fits in FLUX 3
FLUX 3 is BFL's multimodal family announced in July 2026 (images, video, audio and robot actions on one backbone). FLUX 3 Image Klein is the second open-weights release in the family, after the 7B robotics model FLUX 3 Action (September 2026); BFL says more open releases will follow. “Klein” is German for “small” — the name BFL also used for the FLUX.2 [klein] models of January 2026.