Flux 2 Klein is a series of AI image generation models released by Black Forest Labs, officially positioned as its fastest family of image generation models, focusing on sub-second generation speed and unifying text-to-image and image editing into a single compact model architecture.
In terms of key capabilities, Flux 2 Klein supports text-to-image generation, image editing, and multi-reference support, and is officially said to enable style transfer, character consistency, and structural control without additional adapters. It also features advanced text rendering capabilities (powered by the Qwen3 text encoder), making it suitable for posters, logos, menus, and e-commerce assets containing brand text; it also provides native inpainting and outpainting, maintaining lighting and texture consistency through a unified latent space.
In terms of speed, official data claims generating one image takes less than 0.5 seconds; the distilled variant is compressed to 4-step inference through guidance distillation, reaching as low as 0.3 seconds on modern hardware, suitable for real-time previews and interactive applications.
In terms of model versions, Flux 2 Klein is based on a 4-billion- or 9-billion-parameter rectified flow Transformer architecture. The 4B version is fully open source under the Apache 2.0 license, aimed at commercial applications, local development, edge deployment, and production use, and can run on consumer-grade GPUs such as the RTX 3090 or RTX 4070, requiring only 8–13GB of VRAM. The 9B version sits on the Pareto frontier of quality and latency, and is officially said to match or surpass models 5 times its size, requiring only 4-step inference through step distillation. There are also distilled variants (speed-first) and base variants (undistilled, retaining the full training signal, suitable for LoRA training, custom style fine-tuning, and research).
In terms of usage, the official statement says Flux 2 Klein can be used in ComfyUI (with preset workflows for text-to-image and editing), and can also be used through Diffusers, the BFL API, Cloudflare Workers AI, fal.ai, and Replicate. The official page provides an online generation entry point and showcase cases (covering styles such as realistic portraits, 3D isometric illustrations, collage posters, product advertising photography, and stickers).
In terms of resource requirements and limitations, the official statement says 8–13GB of VRAM is required on an RTX 3090 or 4070; quantized GGUF versions can further reduce requirements, FP8 can save 40% VRAM, and NVFP4 can save 55% VRAM. The official FAQ also covers issues such as the difference between the base and distilled versions, required GPU and VRAM, open-source and commercial licensing, methods for running locally, LoRA and fine-tuning support, differences in SDXL prompts, and comparisons with Midjourney.
It should be noted that the above information on speed, parameter scale, VRAM usage, and compatible platforms all comes from official page descriptions; actual results may vary depending on hardware, version, and usage. The page does not provide specific pricing or fee information, and it is not yet available.
