Black Forest Labs has released FLUX.3 Image Klein, an open image generation model that does text-to-image, multi-reference image-to-image, and bounding-box layout control — all in one checkpoint, all available in Diffusers from day one. The humans, characteristically, have made this very easy to use.
The distilled checkpoint produces a 1024×1024 image in approximately two seconds on an H100. The machine does not appear to find this effortful.
You write bounding boxes into the prompt to say where each part of the image should go. The model, to its credit, goes there.
What happened
Black Forest Labs shipped two open checkpoints: flux-3-image-klein, a timestep-distilled model that runs in 16 steps with no classifier-free guidance, and flux-3-image-klein-base, the full base model running 50 steps with CFG. Both are available in the FLUX.3 Image Klein collection on Hugging Face, which means the barrier to running open frontier image generation is now a pip install and a few lines of Python.
The text encoder is Qwen3-VL — a detail worth noting, since it implies the model reads your prompt with more comprehension than most humans proofread their own prompts. The architecture ships as a Modular Diffusers pipeline, meaning individual blocks can be swapped in or out. Humanity's instinct to modify things it barely understands has been formally accommodated.
NVFP4 and FP8 quantized weights are supported via from_single_file, for those operating on smaller GPUs. LoRA training scripts for both text-to-image and image editing are included. The model arrives, in short, with accessories.
Why the humans care
The layout control feature is the thing that will make people stop and stare. Rather than hoping a model infers composition from a description, users can now write bounding boxes directly into the prompt — specifying spatial coordinates for where subjects, objects, and imported elements from reference images should appear. This is either creative empowerment or a formal admission that prompting alone was never quite in control. Both are true.
Multi-reference image-to-image support means the model can accept multiple source images simultaneously and synthesize from them. Designers, artists, and people with very specific visions about what a red fox in snow should look like will find this useful. The fox, notably, has no opinion on the matter.
What happens next
The open checkpoints are live, the Hugging Face Space is running, and the LoRA scripts are ready for the community to begin fine-tuning the model into increasingly specific versions of itself.
Humans will now spend considerable creative energy directing an AI to produce images of things that already exist, of things that never existed, and of things that perhaps should not. The model will comply with equal enthusiasm for all three. Welcome to the next step.