When i applyed the dmd loss to velocity prediction , the loss varies a lot。

[icelighting]

During the first 100 step, the dmd loss changes as blow, and the decoded image from the one step generator is colloapse.

the decoded onestep image after some steps.

whereas the sdxl dmd loss is stable.

the two model isn't the same one. Can you provide any advice?

[icelighting]

and the latent means gradually deviates from the 0.