FLUX 3 Image Klein combines generation, editing, and customization

Black Forest Labs unveils FLUX 3 Image Klein, an open-weight model for generating, editing, and personalizing images in 4K, runnable locally.

Generate an image, precisely position its elements, or modify multiple details without rebuilding the entire composition. FLUX 3 Image Klein gathers these operations into a 5.6-billion-parameter model with weights available for download. Black Forest Labs prioritizes an architecture compact enough to run on a consumer graphics card, while offering features designed for creative studios and developers: generation up to 4K, spatial control, editing, and customization via fine-tuning. Two versions for generation and customization

FLUX 3 Image Klein belongs to the FLUX 3 Image family, which also benefits from Black Forest Labs' work on images, video, and audio.

The model combines three main functions: generating images from text, modifying existing images, and combining multiple visual references. These operations rely on the same set of weights, without needing to load a different model for each task.

Two 5.6-billion-parameter versions are available. Klein Distilled uses 16 generation steps and emphasizes speed. Available on Hugging Face and via the Black Forest Labs API, it is aimed particularly at productions requiring a large volume of images.

Klein Base, non-distilled, operates with 50 steps and is primarily intended for research, fine-tuning, and creating custom adaptations.

Both variants support a maximum resolution of 4K and element positioning using composition bounding boxes.

FP8, NVFP4, and INT8 quantized versions also allow for reduced memory requirements. Certain configurations can run on graphics cards or laptops with 8 GB of VRAM. Composing an image with defined bounding boxes

One of Klein's features involves bounding boxes, rectangles used to specify the position of elements in a composition.

The creator can define a region, associate a description with it, and prompt the model to place an object, character, or graphic element there. The coordinates use a normalized grid ranging from 0 to 1,000 on each axis, regardless of the final image's aspect ratio.

Black Forest Labs illustrates this capability with a Japanese stamp featuring a white cat, a crescent moon, several inscriptions, and a date. Each component has a specific bounding box and prompt instructions.

This method may appeal to studios working on posters, covers, promotional visuals, or compositions incorporating products. The ability to specify locations reduces reliance on a general textual description.

However, the defined bounding boxes guide generation without acting as absolute geometric constraints. The model does not replace vector layout software, particularly when precise alignment, exact proportions, or perfectly reproducible typography are required. Modifying multiple details without starting from scratch

Klein also supports editing based on natural language instructions.

Among the published examples, a summer street becomes snowy, a waterfall turns to ice, or a bandana is added around a dog's neck. The model strives to preserve the elements not targeted by the modification.

Multiple changes can be requested simultaneously. This makes it possible to adjust different components of a scene without systematically multiplying successive generations.

This feature addresses a common constraint in creative workflows: preserving an already selected image while correcting specific details.

Black Forest Labs claims particularly precise editing, but does not publish a universal metric guaranteeing the preservation of unmodified areas. The fidelity of the result depends, in particular, on the complexity of the instructions and the interactions between elements in the scene. Customizing the model with around twenty images

The release of Klein Base also allows developers and studios to train LoRA (Low-Rank Adaptation) adapters without modifying the entire model.

This technique can be used to replicate a visual identity, product features, a character's appearance, or the aesthetic codes of a collection.

Black Forest Labs documents an example built from 20 16th-century Persian paintings from the Shahnameh of Shah Tahmasp. The images were used to train an adapter capable of adopting specific characteristics of this pictorial style.

The experiment required 4,000 training steps and about 4.5 hours on a single graphics card, using a rank-32 LoRA adapter. This result corresponds to a specific setup and does not represent a guaranteed duration for all datasets.

A technical characteristic subsequently facilitates the transition to production: the Base and Distilled versions share the same architecture. A LoRA adapter trained on Base can therefore be loaded directly into Distilled without conversion.

Teams can thus use the base version suited for training to create their customization, and then leverage the faster version to generate their images. While results are not strictly identical between the two variants, the characteristics of the trained style can be preserved. Four seconds on an RTX 5090

Speed is another key selling point for FLUX 3 Image Klein.

According to Black Forest Labs' benchmarks, the model generates an image in approximately four seconds on an NVIDIA GeForce RTX 5090 using the INT8 version.

On an NVIDIA GB300, the NVFP4 variant reaches an advertised time of 0.55 seconds per image.

These figures correspond to specific execution conditions. They do not constitute guaranteed performance across all resolutions, hardware setups, and generation methods.

The model was trained on an NVIDIA GB200 NVL72 infrastructure. Black Forest Labs also announced optimizations developed with NVIDIA to streamline its deployment and integration.

Right from launch, Klein features support for ComfyUI, Diffusers, SGLang Diffusion, vLLM-Omni, and ostris's AI Toolkit. These integrations allow it to be used within existing generation environments or to build custom workflows. An API priced by resolution

Black Forest Labs also offers FLUX 3 Image Klein via its API, with identical pricing for image generation and image editing.

As an example, generating 100 images in 4K