In Pony's training, source tags are associated with the various types of images that the checkpoint learns from during training. (Furry, Pony, Anime, ...)

As a result of such training, these tags become associated to the styles that the model 'sees' in those images. Hence, similarly to prompting source_furry in the negatives when you would like to generate anime art, we put source_pony in there too, to reduce the influence of those styles in our generations.

On the other hand, we may also opt to not prompt source_anime in the prompt as it has similarly learnt a style of anime that may conflict with style LoRAs, leading to a slightly less satisfactory result. Do note that the score tags also hold some inherent style due to biases that may be prevalent during the rating of images.

If we are using Auto1111, we can circumvent the application of the Pony's inherent style from the score tags if we limit their application during generation by a factor of the number of steps while retaining the benefits of better anatomy from the score tags. (Using Prompt Editing)

Score tags are only active for the first 20% of the generation with the below prompt.

[score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up::0.2]

Feel free to drop suggestions.

While I don't take commissions, I might bake a lora if I like the style.