FLUX 3 Image Klein is a 5.6B parameter open-weights model designed for text-to-image generation, image editing, and multi-reference generation (supporting up to 10 references). It serves as the second release in the FLUX 3 family, leveraging cross-modal learning—where spatial details from images, motion dynamics from video, and event information from audio are synthesized to produce images with improved physical grounding.

Core Methodology and Capabilities

The model architecture focuses on high-fidelity visual synthesis, featuring native 4K resolution support and pixel-perfect editing capabilities. A primary technical feature is the integration of bounding box controls, which allow users to specify spatial locations for generated elements or restrict edits to localized regions.

FLUX 3 Image Klein achieves superior performance relative to its parameter size by utilizing a distilled architecture alongside a base model optimized for fine-tuning. The development process benefited from training on NVIDIA GB200 NVL72 hardware and optimization for the NVIDIA accelerated computing stack, including TensorRT-LLM, vLLM-Omni, and SGLang Diffusion.

Technical Performance and Optimization

The model is engineered for high efficiency and local deployment:

• Memory Requirements: The generator weights are 11.2 GB in BF16, 5.6 GB in FP8, and further optimized for INT8/NVFP4 builds. Through INT8 quantization, the model can operate on consumer GPUs with 8GB of VRAM via ComfyUI or Diffusers.
• Latency: Utilizing the step-distilled model (16 steps) and an INT8 build, 1MP images can be generated in approximately 4 seconds on an RTX 5090. On high-end server hardware (NVIDIA GB300), the NVFP4-optimized distilled model achieves latency below 0.6 seconds for 1MP generation.
• Comparative Metrics: Internal evaluations place FLUX 3 Image Klein ahead of larger models (e.g., Qwen-Image 2.1 at 7B and Krea 2 at 12.9B) in text-to-image quality, multi-reference generation, and image editing. Specifically, it provides an Elo gain of 91 in text-to-image and 70 in image editing compared to its predecessor, FLUX.2 [klein], while reducing the parameter count by 38%.

Ecosystem and Deployment

The release includes:

1. Base Model: Designed for LoRA fine-tuning, supported by Ostris's AI Toolkit and NVIDIA NeMo AutoModel.
2. Distilled Model: Optimized for low-latency production workloads.
3. Deployment Support: Native compatibility with Hugging Face Diffusers, ComfyUI, and various inference backends.

Licensing and Safety

The model is governed by the "FLUX Kommunity License," which permits free commercial use for entities with annual revenue under $5M. Safety is addressed through pre-training dataset filtering and post-training targeted fine-tuning. The model underwent rigorous internal and third-party red teaming (via Cinder) to mitigate potential misuse and harmful outputs.