FLUX-Lightning Slashes Diffusion Inference to 4 Steps, Doubling Speed
FLUX-Lightning, introduced by PaddleMIX, combines phased consistency distillation, adversarial learning, distribution‑matching distillation, and reflow loss to reduce diffusion model inference to just four steps while preserving image quality, and leverages the CINN compiler to achieve over 30% speed gains on A800 GPUs, surpassing existing SOTA acceleration methods.
