FLUX.3 Image Klein Brings 2-Second AI Image Generation to Diffusers
Black Forest Labs launches FLUX.3 Image Klein in Diffusers with layout control, multi-reference editing and 1.93-second H100 generation on day one.
Summary
Black Forest Labs released the open FLUX.3 Image Klein on October 15, 2026, with immediate Diffusers support. The black-forest-labs/flux-3-image-klein distilled checkpoint generates 1024 × 1024 images in 16 steps without classifier-free guidance, while black-forest-labs/flux-3-image-klein-base uses 50 steps with guidance and supports fine-tuning. One model handles text-to-image, multiple-reference image editing and layout control. Prompts can place regions through JSON bounding boxes normalized from 0 to 1000, or move reference-image content between source and target boxes while preserving pinned areas.
On one NVIDIA H100 80GB, the distilled model completed text-to-image in 1.93 seconds, single-reference image-to-image in 3.87 seconds and layout editing in 4.10 seconds, each peaking at 23.8 GiB. A 1024 × 1024 reference adds 4,096 tokens per transformer call and roughly doubles runtime. BF16, rowwise FP8 and NVFP4 single-file weights are available; NVFP4 used 13.4 GiB after loading on an RTX6000. GPUs without compatible fast kernels receive lower memory use but no acceleration.
Qwen3-VL supplies text features, while the Flux.2 VAE converts each 1024 × 1024 image into 4,096 tokens. Five separate-stream blocks process text, generated-image and reference tokens before 28 joint transformer blocks merge them. DreamBooth LoRA scripts support generation and instruction-based editing, while Modular Diffusers lets developers replace blocks, cache prompt embeddings and use CPU offloading. Both checkpoints are available through Hugging Face.
Positives
- 1.93-second generation produces a 1024 × 1024 image in 16 steps on one NVIDIA H100 80GB.
- One FLUX.3 Image Klein model supports text-to-image, multiple-reference editing and normalized bounding-box layout control.
- BF16, rowwise FP8 and NVFP4 single-file weights broaden deployment options for GPUs with less memory.
- DreamBooth LoRA scripts support both image generation and instruction-based image editing.
- Modular Diffusers allows block replacement, cached text embeddings and automatic CPU offloading.
Risks & concerns
- 23.8 GiB peak memory across measured H100 workflows remains demanding for consumer hardware.
- One 1024 × 1024 reference adds 4,096 tokens and increases runtime from 1.93 seconds to 3.87 seconds.
- FP8 acceleration requires compute capability 8.9 or newer, covering Ada, Hopper and Blackwell GPUs.
- NVFP4 fast kernels require Blackwell hardware and matching flashinfer-python and flashinfer-cubin versions.
- Unsupported GPUs must dequantize weights during every step, reducing memory use without improving speed.