SANA-Sprint-1.6B

SANA-Sprint-1.6B is an ultra-efficient text-to-image model by NVIDIA Labs (Sana family) with about 1.6 billion parameters. It uses a hybrid distillation strategy combining continuous-time consistency distillation (sCM) with latent adversarial diffusion distillation (LADD), which enables high quality image generation in very few inference steps (1-4 steps). It achieves a strong speed-vs-quality trade-off, making it ideal for real-time visual tasks and interactive generation.