1.58-bit FLUX实现顶级文生图模型的极低比特量化，不依赖图像数据。
1.58-bit FLUX
- 用{-1,0,+1}三值权重对FLUX.1-dev进行1.58比特量化，仅靠模型自监督。
- 存储减7.7倍、推理内存减5.1倍，生成1024×1024图像质量接近原模型。
- 专为三值计算优化内核，适合部署在资源受限设备上。
我们提出1.58-bit FLUX，首次成功将顶级文生图模型FLUX.1-dev以1.58比特权重（即取值为{-1, 0, +1}）进行量化，同时保持生成1024×1024图像的相近性能。该量化方法无需图像数据，仅依赖FLUX.1-dev模型自身的自监督。此外，我们开发了专用于1.58比特运算的定制内核，实现模型存储降低7.7倍、推理内存减少5.1倍，并提升推理延迟。在GenEval和T2I Compbench基准上的大量评估表明，1.58-bit FLUX在显著提升计算效率的同时，仍能维持生成质量。
原文摘要 · Abstract (English)
We present 1.58-bit FLUX, the first successful approach to quantizing the state-of-the-art text-to-image generation model, FLUX.1-dev, using 1.58-bit weights (i.e., values in {-1, 0, +1}) while maintaining comparable performance for generating 1024 x 1024 images. Notably, our quantization method operates without access to image data, relying solely on self-supervision from the FLUX.1-dev model. Additionally, we develop a custom kernel optimized for 1.58-bit operations, achieving a 7.7x reduction in model storage, a 5.1x reduction in inference memory, and improved inference latency. Extensive evaluations on the GenEval and T2I Compbench benchmarks demonstrate the effectiveness of 1.58-bit FLUX in maintaining generation quality while significantly enhancing computational efficiency.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。