What it is
Training-free staged sampler for flow-matching text-to-image models that shifts most denoising to low resolution, then does 1-step high-res refinement for major speedups. Accelerates pretrained flow-matching image generators by sampling at low resolution, upscaling with Real-ESRGAN, re-encoding, adding scheduler-consistent noise, and doing a short high-res refinement.
Key results
- 10.3x end-to-end speedup on Qwen-Image with 12+1 steps
- 4.77s vs 49.32s latency on Qwen-Image 12plus1 vs native 50-step
- Works across Qwen-Image, FLUX.1-dev, FLUX.2 Klein, and Z-Image
Shows that large diffusion speedups are possible by moving most denoising to low resolution and only refining once at high resolution.