1.58-bit FLUX

TL;DR

This paper introduces 1.58-bit FLUX, a quantization method for text-to-image models that reduces storage and computation without sacrificing image quality, using self-supervision and custom hardware optimization.

Contribution

The paper presents the first quantization of FLUX model to 1.58-bit weights without data access, along with a custom kernel for efficient inference.

Findings

7.7x reduction in model storage

5.1x reduction in inference memory

Maintains image quality on benchmarks

Abstract

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.