Black Forest Labs has released FLUX.3 Image Klein, an open image-generation model that combines text-to-image, multi-reference image-to-image, and explicit layout control via bounding boxes. The project was published on October 15, 2026 and ships as a Modular Diffusers pipeline with both a distilled and a base checkpoint.
Capabilities and workflow
The model accepts scene prompts and optional reference images. Layout control is expressed by adding a JSON list of regions after the caption; each region includes an id, a bbox and a description. Bounding boxes are given as [y_min, x_min, y_max, x_max] on a 0–1000 normalized scale so the same boxes work at any output size. For image edits, regions can specify src_bbox (from the reference) and tgt_bbox (target location).
Black Forest Labs provides two checkpoints in the FLUX.3 Image Klein collection: the distilled checkpoint (black-forest-labs/flux-3-image-klein) which runs with 16 steps and no classifier-free guidance by default, and the base checkpoint (black-forest-labs/flux-3-image-klein-base) which runs 50 steps with guidance and is intended for fine-tuning and LoRA training. The team notes the distilled checkpoint can generate a 1024 × 1024 image in about 2 seconds on an NVIDIA H100.
Architecture and performance details
FLUX.3 is a latent-space flow model that uses a VAE (the same VAE as Flux.2) to compress images 8× per axis so a 1024 × 1024 image becomes a 64 × 64 grid, or 4,096 image tokens. The text encoder is Qwen3-VL; FLUX.3 stacks hidden states from eight layers (layers 4, 8, …, 32) into a 20,480-d vector before projecting to the transformer width of 3,072.
The transformer runs in two stages: five per-mode blocks that process text, noisy image and reference streams separately, followed by 28 joint blocks that process the concatenated sequence [text | image | reference]. Black Forest Labs reports token counts per transformer call of 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.
Memory benchmarks published by the developer show peak allocated memory (after load) varying by variant: nvfp4 (RTX6000) 13.4 GiB, fp8r (H100) 14.35 GiB, and bf16 (H100) 18.80 GiB. End-to-end timings for the distilled checkpoint (16 steps) on a single H100 80GB are given as about 1.93 s for 1024 × 1024 text-to-image with a peak memory of 23.8 GiB; a 1024 × 1024 reference roughly doubles runtime because it adds 4,096 tokens to every transformer call.
Open checkpoints, quantized single-file transformer variants (fp8r, nvfp4), and LoRA training scripts for both text-to-image and image-to-image are included in the FLUX.3 Image Klein collection: black-forest-labs/flux-3-image-klein and black-forest-labs/flux-3-image-klein-base. The release notes detail loading, quantization, and training options, plus example prompts and region-format examples.
Original source: Hugging Face Blog