Diffusion models achieve high quality image reproduction but are limited by slow repetitive sampling. Distillation methods reduce this by allowing one or more generations. Flow simulation, which was initially presented as a different framework, has since been shown to be similar to a distribution under the Gaussian assumption, which raises the question of whether the distillation techniques are similar to direct point distillation. We provide a simple derivation – based on Bayes' rule and conditional expectations – that combines Gaussian diffusion and flow coherence without relying on the ODE/SDE formulation. Building on this idea, we extend Score identity Distillation (SiD) to pre-trained models that align text-to-image flows, including SANA, SD3-Medium, SD3.5-Medium/Large, and FLUX.1-dev, all with DiT backbones. Tests show that, with only limited flow- and DiT-specific adjustments, SiD works out of the box for all of these models, in both data-free and data-assisted settings, without requiring teacher adjustments or architecture changes. This provides the first systematic proof that point distillation is widely applicable to text-to-image flow models, resolving previous concerns about stability and noise and integrating speedup techniques across flow-based generators.
- † University of Texas at Austin
