Tongyi Proposes Latent-to-Pixel Strategy to Speed Up Diffusion Models

Alibaba's Tongyi team proposes a latent-to-pixel training strategy that speeds up convergence of pixel-space text-to-image diffusion models and improves inference speed, reportedly making them up to 4.75x faster than latent models.

2026-08-18 ~ 2026-08-20 · 2 related posts