ArcFlow Generates AI Images in Just Two Steps

By vramkickedin | | 2 min read
Italic orange text for Arcflow on a rolling digital hill

ArcFlow is a new framework that generates images from text prompts in just two processing steps. It achieves this by using curved mathematical paths instead of straight shortcuts, which better matches how diffusion models actually work.

Developed by researchers from Fudan University and Microsoft Research Asia, ArcFlow solves a key problem in AI image generation. Standard diffusion models require many sequential steps to create images, making them slow. ArcFlow distills this process while keeping image quality high, working with popular models like FLUX.1-dev and Qwen-Image-20B.

Model Size: Varies by base model & VRAM GPU: 34GB-57GB required

What features ArcFlow offers

  • Creates images in only 2 steps with minimal quality loss.
  • 40x speedup compared to original multi-step models.
  • Requires fine-tuning of less than 5% of original parameters.
  • Works with both FLUX.1-dev and Qwen-Image-20B base models.
  • Includes CPU offload feature to reduce graphics card memory needs.
  • Supports combination with style adapters and other LoRAs.

Creative professionals working with AI image generation will appreciate the faster iteration times. Small studios and individual creators can produce high-quality images without waiting through dozens of processing steps, making workflows more efficient.

Technical requirements and hardware demand

The research team identified that existing distillation methods fail because they use linear approximations that cannot keep up with changing velocities during image generation. ArcFlow solves this by treating the velocity field as a mixture of continuous momentum processes, allowing for

'high-precision approximation of the teacher trajectory.'

Hardware demands are significant. The GitHub repository states that:

'our inference requires ~57GB for ArcFlow-Qwen and ~34GB for ArcFlow-FLUX.'

Users with limited graphics memory can enable CPU offloading to bring requirements down to 41GB and 25GB respectively. The developers recommend keeping the timestep_ratio setting at 1.0 when using the released checkpoints for best results.

Get ArcFlow on GitHub or download the models from Hugging Face. Read the full research in their arXiv paper.