SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

FLUX.2 [klein] - Image Generation Model Open-Sourced by Black Forest Labs

Author: AI小集 · Original source: AI工具集

What is FLUX.2 [klein]

FLUX.2 [klein] is an efficient open-source image generation and editing model from Black Forest Labs. The model features sub-second inference speed, completing high-quality image generation and editing within 0.5 seconds, and supports text-to-image, image-to-image, and multi-reference generation. The 4B version of the model requires only 13GB of VRAM, making it compatible with consumer-grade GPUs such as the RTX 3090/4070, while the 9B version offers stronger performance. The model provides open weights, supports deep customization and fine-tuning, and offers FP8 and NVFP4 quantized versions to further optimize speed and VRAM usage.图像生成与编辑模型。模型具备亚秒级推理速度，能在 0.5 秒内完成高质量图像生成和编辑，支持文生图、图生图及多参考生成。模型 4B 版本仅需 13GB 显存，适配消费级 GPU，如 RTX 3090/4070， 9B 版本性能更强。模型提供开放权重，支持深度定制与微调，推出 FP8 和 NVFP4 量化版本，进一步优化速度和显存占用。

FLUX.2 [klein]

Main Features of FLUX.2 [klein]

  • Sub-second inference: On modern hardware, image generation or editing can be completed within 0.5 seconds, suitable for real-time applications.
  • High-quality output: The images generated by the model have high realism and diversity, with outstanding performance in the base version.
  • Unified generation and editing capabilities: A single model supports text-to-image (T2I), image-to-image (I2I), and multi-reference generation, meeting a variety of needs.
  • Low hardware requirements: The 4B version of the model requires only 13GB of VRAM and can run on consumer-grade GPUs (such as RTX 3090/4070), making it easy to deploy.
  • Quantization optimization: The model offers FP8 and NVFP4 quantized versions, further improving speed and reducing VRAM usage.

Technical Principles of FLUX.2 [klein]

  • Compact integrated architecture: The model unifies image generation and editing capabilities in a single model, achieving low-latency inference through efficient architecture design.
  • Step distillation technology: By compressing the inference steps to 4, inference speed is significantly improved while maintaining high-quality output.
  • Flow model and text embedder: Built on a 9B Flow model and equipped with an 8B Qwen3 text embedder, it achieves powerful text understanding and image generation capabilities.
  • Quantization optimization: The FP8 and NVFP4 versions developed in collaboration with NVIDIA further reduce VRAM usage and improve speed by optimizing data precision and computational efficiency.
  • Base model and distilled model: The model offers an undistilled base model (retaining all training signals, suitable for fine-tuning and customization) and a distilled model (optimized for speed and latency).

Project Links for FLUX.2 [klein]

  • Project official website：https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence
  • HuggingFace model repository：
    • https://huggingface.co/spaces/black-forest-labs/FLUX.2-klein-9B
    • https://huggingface.co/spaces/black-forest-labs/FLUX.2-klein-4B

Application Scenarios of FLUX.2 [klein]

  • Real-time design tools: FLUX.2 [klein] can be used in scenarios such as advertising, UI/UX design, and game development. Its sub-second generation speed and high-quality output enable rapid iteration of design solutions and improve design efficiency.
  • Content creation and editing: The model is suitable for social media content generation, video editing, and animation production, quickly generating diverse content and improving creative efficiency.
  • AI-driven interactive systems: The model is suitable for scenarios such as AI assistants, intelligent customer service, and virtual avatar generation. With real-time response capabilities, it enables interactive visual dialogue and generates images that meet user needs.
  • Education and training: In education, virtual laboratories, and vocational training, FLUX.2 [klein] can quickly generate high-quality educational materials, such as diagrams and simulated scenes, enhancing the learning experience.
  • Game development: The model is suitable for in-game scene generation, character design, and dynamic backgrounds, etc. Its low hardware requirements enable it to run on consumer-grade devices, supporting real-time rendering and content generation in game development.