Black Forest Labs’ FLUX.3 Image Klein Lands in Hugging Face Diffusers, With Layout Control Built In

Black Forest Labs’ new open image model, FLUX.3 Image Klein, runs in Hugging Face’s Diffusers library from launch day. A single model handles text-to-image generation, editing from one or more reference images, and a new layout feature that lets users say where each part of the picture should go, according to a Hugging Face blog post published on 15 October 2026.

Diffusers is Hugging Face’s widely used open-source Python library for running image-generation models. The post was written by Hugging Face’s Aritra Roy Gosthipaty and Sayak Paul.

Two open checkpoints

Hugging Face says two open checkpoints are available in a FLUX.3 Image Klein collection:

• black-forest-labs/flux-3-image-klein, a “time step distilled” version for fast generation, which runs 16 steps by default without classifier-free guidance (CFG);
• black-forest-labs/flux-3-image-klein-base, the base model, which runs 50 steps with CFG and is the one Hugging Face recommends for fine-tuning.

Distillation trains a model to reach a finished image in fewer steps. CFG is a technique that steers the image more closely to the prompt at the cost of extra computation.

What it can do

• Text-to-image. Hugging Face says the model “responds best to a full scene description with the subject, the setting, the light, the composition, and the style.”
• Image-to-image. Users pass one or more reference images and write the prompt as a caption of the image they want. Multiple references are named in the prompt as <ref_image_0>, <ref_image_1> and so on, so one can say, for example, put the fox from one image on the bench from another.
• Layout control. Users add a list of regions after the caption, each with a bounding box and a description. Boxes use a 0 to 1,000 scale, so “the same box works at any image size.” For edits, each region can be moved from a source box in the reference to a target box in the output, while regions with matching boxes stay put.

Speed and memory

Hugging Face’s own tests, on a single NVIDIA H100 80GB, found the distilled checkpoint made a 1024 × 1024 image from text in 1.93 seconds. Image-to-image with one reference took 3.87 seconds, and a layout edit took 4.10 seconds, with peak memory of 23.8 GiB in each case. Hugging Face says a reference image “roughly doubles the time” because it adds about 4,096 tokens to every step.

For smaller GPUs, Black Forest Labs also ships the core model in compressed FP8 and NVFP4 formats. Hugging Face measured peak memory after inference of 15.3 GiB for NVFP4 on an RTX 6000 and 16.27 GiB for FP8 on an H100, against 20.72 GiB for the full bf16 version. The faster FP8 kernels need newer NVIDIA GPUs (Ada, Hopper or Blackwell), and NVFP4 needs Blackwell; on other GPUs the weights save memory but give no speed-up.

Under the hood

The text encoder is Qwen3-VL, a vision-language model. FLUX.3 reads eight of its 36 layers, not just the last, to give the image model a richer view of the prompt. The image model processes text, the image being generated and any reference images first in five separate blocks each, then in 28 shared blocks where everything can attend to everything else. It reuses the FLUX.2 image encoder (VAE).

Diffusers ships the model as a “Modular Diffusers” pipeline, so developers can remove steps, such as running the text encoder once and freeing its memory, or add their own. It also includes DreamBooth LoRA training scripts for both text-to-image and image-editing fine-tunes.

What’s still unknown

The Hugging Face post does not set out the licence terms, and TSN has not reviewed the model cards. Speed and memory figures are Hugging Face’s own benchmarks on specific hardware, and no independent quality comparisons have been published yet.

What to watch

Watch for Black Forest Labs’ licence details, community LoRAs built on the base checkpoint, and independent tests of the layout feature.

What this means

• One model, three jobs. Generation, multi-image editing and layout control in a single open model.
• Fast on top hardware. About 2 seconds per image on an H100, by Hugging Face’s figures.
• Smaller formats available. FP8 and NVFP4 versions cut memory use.

Sources

1. Aritra Roy Gosthipaty and Sayak Paul, “Welcome FLUX.3 Image Klein in Diffusers: The latest open image generation model from Black Forest Labs”, Hugging Face blog, 15 October 2026

Related stories

• Sperid Labs Releases Iris-3B, an Open Image Model With No VAE; Its Own Paper Finds No Clear Gain for Depth or Restoration
• NVIDIA Releases Nemotron 3.5 Super VL, an Open 120B Multimodal MoE for Agents