FLUX 3 Image Klein opens weights for reference-led image editing
Black Forest Labs released FLUX 3 Image Klein for image generation and editing, including work guided by up to ten reference images. Bounding boxes specify where elements belong or which region should change. Base and distilled models offer adaptation and faster-generation paths. Local deployment tools accompany the weights, whose use remains subject to the FLUX Kommunity License.
Artificial Intelligence··Night
Klein combines generation with image editing
Black Forest Labs, the developer of the FLUX image-model family, released FLUX 3 Image Klein on October 15. The model supports generating images from text and editing existing pictures. The launch includes a local path through Diffusers, a software library for image-generation models, with controls for guiding the placement of objects.[1], [2]
The generator has 5.6 billion parameters, excluding its text encoders. It can use up to ten reference images in one generation workflow. Bounding boxes describe where an element should appear or which part of an existing image should be edited. A localized editing path is intended to preserve pixels outside the chosen area, alongside native 4K output.[1]
Base and distilled models serve different workflows
The base checkpoint is available for adaptation, while a step-distilled version targets lower-latency image generation. NVIDIA-supported FP8, INT8 and NVFP4 builds offer different numerical precision choices. Local operation is documented through ComfyUI, a visual workflow tool, as well as Diffusers. This gives developers a way to work with the model on their own infrastructure.[1]
Adaptation tools accompany licensed weights
For adapting the base model, the release supports LoRA training, a method for adjusting a model through a smaller set of additional parameters. The documented routes include Diffusers and Ostris’s AI Toolkit. NVIDIA NeMo AutoModel also supplies training recipes. Weights and API access are available with the launch, with distribution governed by the FLUX Kommunity License.[1]
Black Forest Labs’ comparative quality and speed results are internal evaluations. The company distinguishes GPU computation time from hosted latency, which can also include queuing, network delays and prompt expansion. Its safety work includes training-data filters, post-training interventions and evaluation with Cinder. Open weights make the model accessible for deployment while its license continues to govern use.[1]
