Page 1 of 1
Phaze-A versus legacy models
Posted: Wed Feb 15, 2023 6:14 pm
by MattB
It occurred to me today that I'm training using Dfl-128 (the "native" model) which has virtually no settings. But should I be using Phaze-A with the "model_phaze_a_dfl-h128_preset" instead? I don't know if the two are genuinely equivalent, if the latter is approximating the former, or if all models should be trained using Phaze-A now? I recently used the dny-512 preset with very good results. In fact, I don't even know if what I'm thinking of as "legacy" models are truly legacy or "as good or better" than Phaze-A.
More informed and capable opinions appreciated.
Re: Phaze-A versus legacy models
Posted: Wed Feb 15, 2023 10:42 pm
by torzdf
The Phaze-A variants are exact replicas of the original models. Theoretically you could load the weights from a standalone model into the Phaze-A version (you can't in practice due to slightly different ways in which the models are constructed). They are, to all intents and purposes, byte compatible.
They were added to Phaze-A because 1) I could and 2) I thought it was useful to have some presets in there structured as existing models.
Ultimately, use what is best for you. I never recommend people start with Phaze-A because, frankly, its overwhelming. However, if you want to start tweaking models yourself, then Phaze-A is the way to go.