FLUX.1 Image Synthesis

Transformer-powered Flow Models at Scale

All public FLUX.1 models are built on a hybrid architecture that combines multimodal and parallel diffusion transformer blocks, scaled to an impressive 12 billion parameters. These models represent an advancement over previous diffusion models by incorporating flow matching, a general and straightforward method for training generative models, with diffusion as a special case.

The FLUX.1 models also enhance performance and hardware efficiency by integrating rotary positional embeddings and parallel attention layers. This combination of technologies ensures that FLUX.1 not only delivers superior image synthesis but also does so more efficiently than its predecessors. A detailed technical report on these innovations will be published soon.

A New Benchmark for Image Synthesis

FLUX.1 establishes a new benchmark in the field of image synthesis, setting higher standards within its model class. The FLUX.1 [pro] and [dev] models surpass popular alternatives like Midjourney v6.0, DALL·E 3 (HD), and SD3-Ultra in various aspects, including visual quality, prompt following, size and aspect ratio variability, typography, and output diversity.

FLUX.1 [schnell], the most advanced few-step model available, outperforms both its in-class competitors and powerful non-distilled models like Midjourney v6.0 and DALL·E 3 (HD). These models are meticulously fine-tuned to preserve the full output diversity achieved during pretraining, offering significantly improved capabilities compared to the current state-of-the-art.

FLUX.1 represents a significant leap forward in text-to-image synthesis, providing a suite of models that cater to both high-end commercial applications and personal use. With its cutting-edge architecture and superior performance across various metrics, FLUX.1 is poised to redefine the standards of image generation.