Tropius LoRA trained on Pony upon request. This one came out shockingly good considering the lack of quality, quantity and character trait consistency in the dataset (nobody could agree on how many wings Tropius is supposed to have or where to put them -_-). There are a variety of issues with this LoRA, but great results are possible and don't require too many attempts.

Issues:

• Has a lot of trouble getting consistent character detail

  • Use reinforcement tags to address this: leaf wings, dark brown skin

• Can add a tail (which Tropius isn't supposed to have) in certain poses

  • Add (tailless, no tail) to your positive and (tail) to your negative

• Wing placement and count can be inconsistent

• Rear views are especially weak with consistency going out the window, be prepared to gen multiple attempts when prompting for rear views

• Clothing tends to come out green if you don't prompt for a color

• Very bad at doing complex poses or situations that veer too far from the source data (which is primarily portraits)

• Has a large effect on the style and content of images, adding greenery to your backgrounds unprompted. This is a side effect of the super low number of training images (<40). The adaptive optimizer had to crank up the strength of the training to compensate for the lack of source data, which resulted in a lot of untagged or poorly tagged background detail getting baked into the base style of the LoRA. (At least, that's my best guess)

Usage Tips:

• If you use 'furry' in your prompting, replace it with 'scalie' for more consistent results

• Use the reinforcement tags listed above to improve the consistency of your gens