The development
What happened
Black Forest Labs’ FLUX.3 Image Klein supports image creation, reference-based editing and explicit composition controls in one Diffusers workflow. Hugging Face’s 15 October guide documents inference, quantization and LoRA customization.
Key facts
Bounding boxes specify where content belongs; editing can map source regions into new target regions.
The distilled checkpoint defaults to 16 steps; the base model uses 50 steps with guidance and is the fine-tuning starting point.
Hugging Face measured 1.93 seconds for 1024×1024 text-to-image and 3.87 seconds for one-reference editing on an H100 80GB, averaged over three warmed-up runs.
BF16, rowwise FP8 and NVFP4 transformer formats are documented. Older GPUs may save memory without faster inference; CPU offloading was not benchmarked.
The context
Why it matters
TideReach assessment: Explicit positioning and reference reuse offer more controllable production workflows than repeated free-form prompting alone. The published H100 measurements help size one deployment, but do not predict consumer-GPU speed or editing reliability across tasks.