When we look at the formulation by flow-matching or DDPM, these are the same thing. We can reparameterise DDPM models to output velocity field as seen in piecewise rectified flow which relies on velocity field as a unified term for distillation.
have you tried similar, or found a reason why this would not work to apply DCM to eg. PixArt Sigma (a DDPM parameterised DiT) or SDXL (an epsilon predicting Unet)?
When we look at the formulation by flow-matching or DDPM, these are the same thing. We can reparameterise DDPM models to output velocity field as seen in piecewise rectified flow which relies on velocity field as a unified term for distillation.
have you tried similar, or found a reason why this would not work to apply DCM to eg. PixArt Sigma (a DDPM parameterised DiT) or SDXL (an epsilon predicting Unet)?