FLUX.3 Image Klein Lets You Draw a Box and Tell the Model Exactly Where to Put Things

Key points

• FLUX.3 Image Klein, released by Black Forest Labs, is the first model in the FLUX.3 family to combine text-to-image generation, multi-reference image editing, and bounding-box layout control in one checkpoint.
• The distilled checkpoint generates a 1024 × 1024 pixel image in roughly 2 seconds on an H100, a high-end Nvidia data-centre GPU.
• Both the full base checkpoint and the distilled version are publicly available on Hugging Face, the platform where researchers share AI models.
• NVFP4 and FP8 quantised weights, compressed versions of the model that use less memory, let it run on smaller GPUs.
• LoRA training scripts, which let anyone fine-tune the model on their own images without retraining from scratch, ship on day one.

Most AI image tools ask you to describe what you want and then hope the model places it somewhere useful. FLUX.3 Image Klein, launched this week by Black Forest Labs and integrated into Hugging Face's Diffusers library on the same day, lets you skip the hope entirely.

You write bounding boxes, simple coordinate rectangles that mark regions of the image, directly into the prompt. One box might say "fill this top-left quarter with sky", another says "put the subject here". The model draws each element inside its assigned region.

What can it actually do?

One model handles three tasks. Standard text-to-image: describe a scene, get a picture. Image-to-image editing using one or more reference images, where you can instruct the model to move a specific object from one part of a photo to another while leaving everything else untouched. And the bounding-box layout control described above.

The layout coordinate system runs from 0 to 1000 on both axes, normalised so the same box description works at any image size. A layout designed for a square image doesn't need rewriting when you switch to a wide crop.

The distilled checkpoint produces a 1024-pixel image in 16 steps and roughly 2 seconds on an H100. The base checkpoint runs 50 steps and is intended for fine-tuning.

For teams without access to top-tier server hardware, Black Forest Labs ships NVFP4 and FP8 weight files that cut memory use, letting the model run on smaller GPUs without a separate download of full-precision components.

Why does the Diffusers integration matter?

Diffusers is the most widely used open Python library for running image-generation models. Shipping on day one means any developer who already uses the library can drop FLUX.3 in with a few lines of code rather than building custom loading logic.

The model arrives as a Modular Diffusers pipeline, a design that lets developers swap individual processing blocks in or out, or add their own. That's less about convenience and more about longevity: a modular architecture means the community can extend the model without waiting for Black Forest Labs to release an update. AI2Day first covered Diffusers on 23 July 2026, when Hugging Face folded compressed model support into the library with a single line of code.

For ordinary people, the most immediate implication is better creative tools. Graphic designers and small studios using image AI will eventually see this layout control reach consumer apps built on open models. Knowing where things will appear before you generate is the difference between a useful tool and an expensive lottery. The big caveat worth watching: every capability described here comes from Black Forest Labs itself, and independent benchmarks haven't yet confirmed how well the layout control holds up on complex scenes.

The open checkpoints are available now in the FLUX.3 Image Klein collection on Hugging Face, and a live demo space is already running on the platform.

Common questions

Do I need expensive hardware to try this?

Not necessarily. The NVFP4 and FP8 quantised weight files are designed for smaller GPUs with less memory. Full performance requires a high-end card, but the compressed versions bring the bar down meaningfully.

What is bounding-box layout control and why does it matter?

A bounding box is a rectangle you define by coordinates, telling the model which part of the image a specific element should occupy. It matters because most image models place objects wherever they see fit; this feature gives the user direct control over composition before a single pixel is generated.

Can I fine-tune this model on my own images?

Yes. LoRA training scripts for text-to-image and image-editing tasks ship with the release. LoRA (Low-Rank Adaptation) lets you customise a model on a small set of your own images without the cost of full retraining.