Black Forest Labs has released FLUX 3 Image Klein, an open-weight image model that can generate pictures, edit existing ones, and combine up to ten reference images. Announced on October 15, it supports output up to 4K and adds bounding-box controls for arranging elements inside an image.
The release includes a faster distilled checkpoint for generation and editing, plus a base checkpoint aimed at fine-tuning. Black Forest Labs also lists support for ComfyUI and Hugging Face Diffusers, giving local users several ways to run it.
More Control Over What Goes Where
According to the official model card, Klein handles text prompts and reference images through the same model. Bounding boxes guide where objects appear and how much space they take up. They can also describe moving, replacing, or removing elements during an edit.
There is a limit to that control. The boxes are guidance, not hard masks. Objects can extend beyond them, and edits can affect areas outside the requested region.
The advertised 5.6 billion parameters refer to the image transformer. The full pipeline also uses a separate Qwen3-VL-4B-Instruct text encoder and an image autoencoder, so the headline model size does not describe the entire installation.
The 8GB Setup Uses a Specific Build
Black Forest Labs’ announcement says the INT8 build can run on NVIDIA GPUs with as little as 8GB of VRAM. The official repository points RTX 3070-class users toward the ComfyUI INT8 files.
That is different from loading the reference implementation. Its published tests on a B200 at 1024 × 1024 report 31GB peak GPU memory in BF16, 26GB in FP8, and 25GB in NVFP4 for the distilled model, with prompt upsampling disabled.
In NVIDIA’s tests cited by Black Forest Labs, the distilled INT8 version generated a one-megapixel image in about four seconds on an RTX 5090 and under nine seconds on an RTX 4070 through ComfyUI. Those are vendor measurements for that setup, not independent results or 4K generation times.
Downloads, Prompting, and License Terms
The reference repository recommends the distilled version for normal generation and editing. It defaults to 16 steps, while the base version uses 50 steps and targets training and research.
It also recommends expanding short requests into detailed scene descriptions before generation. The included prompt upsampler does this automatically, but adds another model to the workflow. Users with limited GPU memory can run that stage separately.
The weights use the FLUX Kommunity License. The model card describes noncommercial use and a commercial-output exception for qualifying users with less than $5 million in gross annualized revenue, including affiliates. That exception does not provide blanket permission for commercial fine-tuning or distribution.
The weights and reference code are available now. For an 8GB GPU, start with the documented ComfyUI INT8 workflow.
