>>5856

The v3 had unintentionally become kind of an 'in between' - version. Entropy and I have been further refining the datasets. Currently a newer version is beeing trained on (Which should be 3 separate Loras for 3 use-cases). V4 and V5 where kind of failures, at least not working as intended. For instance the V5 for 'pole from between breasts' barely works except for when used with one certain Illu model, allthough it was trained on pony.

One of the main problems we've encountered is the fact, that there is a substantial amount of "prior knowledge" about triggers like especially "impaled" in the pony base model. Though it is usually rather hard to pull from that by prompting alone (unless you crank up the cfg) it seemingly conflicts with/or at least negatively influences the Lora training. Because of that we have opted to divide the concept into 2 main categories that are further divided into concepts (~Loras), one for vertical (which is the focus for now/what is beeing trained on) and horizontal. This will be especially necessary for horizontal stuff, because the basemodel will allways try to do something vertical, if the trigger "impaled" is involved. So instead we'll most likely train horizontal stuff on the trigger "skewered".

I have just started to work on refining/reworking the datasest for horizontal stuff again yesterday, so I'm not sure yet if we will also further subdivide horizontal impalement and if so into what exactly. But for vertical impalement there will be one Lora for just the insertion without the pole piercing through and one version each for pole from mouth- and pole from between breasts stuff.

Once the training is done and we've had a few days to test the Loras and confirm they're working they'll be released in Entropy's thread.