That approach just adds a loss from a GAN-type discriminator to the distillation loss when training the distilled sampler. Coalescer models should be able to do generation in at least as few steps as consistency models, so, if you want, you could do the same thing with coalescer models: initial training as normal, then more training that adds in a loss from a discriminator.
That approach just adds a loss from a GAN-type discriminator to the distillation loss when training the distilled sampler. Coalescer models should be able to do generation in at least as few steps as consistency models, so, if you want, you could do the same thing with coalescer models: initial training as normal, then more training that adds in a loss from a discriminator.