FLUX.3 Image Klein, an open image model from Black Forest Labs, accepts layout instructions alongside descriptions. A Hugging Face Diffusers article shows prompts that append bounding boxes on a 0 to 1000 grid, each tied to a description, and edits that move content from a source box in a reference image to a target box. Source.
The release has two checkpoints: a distillationTraining a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.Full definition → one that uses 16 steps by default for faster generation, and a 50-step base model recommended for fine-tuning. The examples show what the controls are for, but they do not establish that every requested object or edit will be rendered faithfully.
Key takeaways · AI-distilled
FLUX.3 layout boxes use y-first coordinates on a normalized 0 to 1000 scale. Editing pairs a source box with a target box; identical boxes specify regions that should stay in place.
FLUX.3's distilled checkpoint defaults to 16 steps without classifier-free guidance. Its base checkpoint uses 50 steps with guidance and is the recommended starting point for fine-tuning.
FLUX.3 quantizationShrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.Full definition → reduces memory use, but speedups depend on hardware and kernels. Unsupported GPUs dequantize weights each step, so lower memory does not imply faster generation.
FLUX.3's reported 1.93-second text-to-image timing used one H100 80GB, 16 steps, no offloading and three runs after warmup. Adding a 1024-pixel-square reference increased the measured time to 3.87 seconds.
FLUX.3's modular pipeline can encode a prompt once, release the text encoder, and reuse its embeddingA list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.Full definition → across multiple seeds. Fewer components need to remain loaded during generation.
Terms in this piece · Glossary
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
LoRA — A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
