FLUX.3 Image Klein Launches: Layout-Controlled Generation Comes to Diffusers

Black Forest Labs announced its FLUX.3 Image Klein image-generation model on the Hugging Face Blog on 15 October 2026; the model also became available in Diffusers that day. FLUX.3 Image Klein combines text-to-image generation and reference-image editing in a single model, while adding layout controls to specify where elements should appear in a scene. The announcement includes two open checkpoints: a distilled version for fast inference and a base model.

How does layout control work?

Users add regions in JSON format to the image description. Each region has an ID, a bounding box and a description; the prompt refers to the relevant element by that ID. The boxes use a normalized scale from 0 to 1000, with vertical and horizontal coordinates in that order. This makes it possible to specify more precisely how much of the image a region should occupy.

For reference-image editing, the system can take separate boxes for an element’s position in the source image and its target position in the output. In the example in the announcement, a character is moved while the background and other objects stay in place. This approach makes it possible to adjust specific parts of the composition rather than only changing the scene description; however, the source does not guarantee that elements will remain unchanged in every case.

Two checkpoints and hardware options

The distilled black-forest-labs/flux-3-image-klein checkpoint runs at 16 steps by default. The base black-forest-labs/flux-3-image-klein-base version uses 50 steps and is intended for fine-tuning. LoRA training scripts for FLUX.3 text-to-image generation and image-editing tasks have also been shared. The training example in the announcement uses the base checkpoint.

In Hugging Face’s benchmark, the distilled version completed 1024 × 1024 text-to-image generation in an average of 1.93 seconds across three runs on a single 80 GB H100. In the same test, editing with one reference image took 3.87 seconds, while a layout edit using a 1360 × 768 reference took 4.10 seconds. The benchmark was run on an H100 with the default attention backend; results may differ on other hardware.

FP8 and NVFP4 quantized weight options are also available to reduce memory use. According to the announcement, FP8r targets GPUs with compute capability 8.9 or higher for fast-kernel support; on other GPUs, weights are dequantized at every step, reducing memory use but not expected to improve speed. NVFP4 requires Blackwell GPUs, along with flashinfer-python and flashinfer-cubin at the same version.

Implications for architecture and visualization workflows

Architects, interior designers and archviz teams can try layout controls to specify where furniture, people or focal elements appear in the frame of concept images. The ability to use multiple reference images also offers an option for bringing elements from different sources together in a single scene. These are potential use cases based on the model’s image-generation and editing capabilities; the source does not announce direct integration with any 3D software.

Teams considering local use should assess GPU memory requirements and the hardware conditions for quantized versions. The announcement does not include licensing or pricing information; licensing details for the checkpoints should be checked separately before commercial production. Model outputs are not described as a delivery format that replaces an existing 3D scene or project file.