Black Forest Labs Releases FLUX.3 Image Klein with Layout Control and Multi-Reference Support

Black Forest Labs has released FLUX.3 Image Klein, a new image generation model now available in Hugging Face Diffusers. The model supports text-to-image, multi-reference image-to-image, and bounding-box layout control, allowing users to specify element positions or move parts of reference images using JSON-defined regions.

Performance and Efficiency

The release includes two checkpoints:

• black-forest-labs/flux-3-image-klein: A time-step distilled model optimized for speed, requiring only 16 steps and no CFG. It generates 1024x1024 images in approximately 2 seconds on an H100 GPU.
• black-forest-labs/flux-3-image-klein-base: The base model requiring 50 steps and CFG, suitable for fine-tuning.

Memory usage is optimized via NVFP4 and FP8 (fp8r) weights. On an H100, the distilled model peaks at 23.8 GiB memory usage. Reference images add 4,096 tokens per transformer call, roughly doubling inference time compared to text-only generation.

Architecture and Features

FLUX.3 uses a Qwen3-VL text encoder and a two-stage transformer architecture:

1. Stage 1 (Per-mode blocks): Separate weights for text, noisy image, and reference streams.
2. Stage 2 (Joint blocks): Shared weights where all tokens attend to each other.

Key features include:

• Layout Control: Define regions using [y_min, x_min, y_max, x_max] normalized to 0–1000. Supports moving elements from src_bbox to tgt_bbox.
• Modular Diffusers: The pipeline allows block replacement and custom insertion, such as Flux3PreviewStep for intermediate results.
• LoRA Training: DreamBooth scripts are provided for both text-to-image and image editing, with support for memory optimization flags like --gradient_checkpointing and --use_8bit_adam.