Black Forest Labs releases FLUX.3 Klein: 1024×1024 in 2 seconds on an H100

Black Forest Labs has released FLUX.3 Image Klein, a new open image generation model. Hugging Face announced Diffusers support on release day. A single model covers text-to-image and image-to-image with multiple references. The main new feature is layout control through bounding boxes written directly in the prompt. The distilled checkpoint produces a 1024×1024 frame in about 2 seconds on an NVIDIA H100.

Frame layout in the prompt

You add a JSON list of regions to the scene description. Each region has an identifier, a bbox and a desc, and the main text refers to the region with a tag such as `<silhouette_1>`. Coordinates are written in the order [y_min, x_min, y_max, x_max] on a normalized 0–1000 scale, so the same box works at any resolution.

For editing, the scheme is extended. Each region gets a src_bbox (where to take the content from in the reference) and a tgt_bbox (where to place it). If the boxes match, the area stays in place. In the developers' example, they move a knitted knight up and to the left while pinning the background, cliff, spire, dragon and fire. When a prompt contains several references, they are called `<ref_image_0>`, `<ref_image_1>` and so on, in the order they are passed.

Speed and memory: what the measurements show

The measurements were made on a single H100 80GB without offloading, averaged over three runs after warm-up, for the distilled checkpoint at 16 steps:

• Text-to-image, 1024×1024: 1.93 s, peak memory 23.8 GiB
• Image-to-image with one 1024×1024 reference: 3.87 s, 23.8 GiB
• Layout editing (the knight), 1360×768 reference: 4.10 s, 23.8 GiB

A reference doubles the time because of the architecture. The Flux.2 VAE compresses the image by 8x along each side into 32 channels, then packs each 2×2 patch into a token with 128 channels. One token covers 16×16 pixels, so a 1024×1024 image yields 4,096 tokens. A reference adds the same number to every transformer call: text-to-image — 4,256 tokens, image-to-image — 8,272, layout text-to-image — 4,336, layout image-to-image — 8,800.

The text encoder is Qwen3-VL. The model takes the hidden states of eight of its 36 layers (4, 8, … 32) and concatenates them: each token becomes a vector of 8 × 2560 = 20,480 dimensions, which a linear layer projects to the transformer width of 3072.

Checkpoints, quantization and fine-tuning

The collection includes two open checkpoints:

• black-forest-labs/flux-3-image-klein — step-distilled, 16 steps, no CFG
• black-forest-labs/flux-3-image-klein-base — base model, 50 steps, CFG; this is the one recommended for fine-tuning

The transformer is available as a single file in bf16 and in two quantized formats. fp8r (rowwise FP8) requires compute capability 8.9+ (Ada, Hopper, Blackwell). On other GPUs the weights are dequantized at every step: this saves memory but gives no speedup. nvfp4 provides fast kernels on Blackwell and requires flashinfer-python and flashinfer-cubin of the same version. Loading goes through from_single_file with BFLQuantizationConfig.

Peak memory after loading and after inference:

• nvfp4 (RTX 6000): 13.4 and 15.3 GiB
• fp8r (H100): 14.35 and 16.27 GiB
• bf16 (H100): 18.80 and 20.72 GiB

The pipeline ships as Modular Diffusers: you can remove blocks or add your own. For tight memory, there is the quantized transformer, a separate text-encoder run that is then freed (it is needed only once per image), and automatic CPU offloading through ComponentsManager (the authors did not benchmark it).

For LoRA, diffusers has two DreamBooth scripts: train_dreambooth_lora_flux3_klein.py for generation and train_dreambooth_lora_flux3_klein_img2img.py for editing on triplets (reference, target, instruction). In the example using the 1920-raider-waite-tarot-public-domain dataset, a tarot style is learned with the trigger trtcrd: rank 16, lora_alpha 16, lr 1e-4, 1500 steps, bf16, batch size 1. The flags --gradient_checkpointing, --cache_latents and --use_8bit_adam save memory. You can also train with Ostris's AI Toolkit.

Where to start

Install the latest diffusers and transformers (`pip install -U diffusers transformers`), load a ModularPipeline from black-forest-labs/flux-3-image-klein in bfloat16, and call the pipeline with prompt and output="images". A detailed scene description covering subject, setting, lighting, composition and style works better than a short phrase. For editing, write the prompt as a caption of the desired result, including the change. On smaller GPUs, use fp8r or nvfp4. You can also try the model in a Space on Hugging Face without installing anything.