Black Forest Labs' distilled FLUX.3 Image Klein renders in about 2 seconds on an H100 in Diffusers

Diffusers support is not the whole install ^[1]. The install line also upgrades transformers, because the text encoder is Qwen3-VL ^[6]. A deployment therefore loads that encoder next to the image model. The load path is three calls on a ModularPipeline: from_pretrained with "black-forest-labs/flux-3-image-klein", load_components with dtype=torch.bfloat16, then .to("cuda") ^[7]. Because the pipeline is modular, you can pop blocks out or add your own ^[8].

Test the 2-second figure first. Divide it by the 16 default steps and each step costs at most about 125 ms ^[18]. It is "at most" because any text encoding or image decoding inside the timing comes out of the same 2 seconds. The post does not say whether the timing includes them or which precision it used. For the number to transfer, your card has to match an H100, the pipeline has to be warm, and you have to run the same 1024 x 1024 at 16 steps ^[13][15].

The RTX 4070 timing is a separate measurement. The Decoder reports under 9 seconds there, on an INT8 build ^[2]. Hugging Face's smaller-GPU weights use other formats. Two publishers, two GPUs and different weight formats mean the gap between 2 and 9 seconds is not a speedup ratio.

The quality claim is Black Forest Labs' own. According to The Decoder, the company says the model scores the highest Elo among open models in internal testing and beats Ideogram V4.5 and FLUX.2 [pro] ^[21]. The same passage lists a 91-point Elo gain over FLUX.2 [klein] ^[22]. An Elo from an internal prompt set measures preference on that set. It transfers to your workload only if those prompts look like yours.

The API price needs a unit check. The Decoder gives $0.009 per 1,000 images ^[17]. That works out to $0.000009 an image ^[19], and a thousand images for under a cent is a number that deserves a second look. Confirm it on BFL's pricing page before it goes into a build-versus-buy sheet.

Layout control has one detail to test early. Boxes are written [y_min, x_min, y_max, x_max], y first, on a 0 to 1000 scale ^[9]. The scale is normalized, so the same box works at any image size. If your boxes come from another tool, check the axis order before anything else. For edits, a region's src_bbox marks where its content sits in the reference and tgt_bbox where it lands in the output. A region with identical boxes stays put ^[10].