It's quiet easy: Start with the prompts, I used, with the same checkpoint and only using this LoRA , no other LoRAs, no Embeddings. This should produce the same results.

After this step, try using an other Checkpoint. Please Note, some Checkpoints are moved very far from the Pony-Diffusion Base-checkpoint. They work less reliable with a LoRA trained with the Base-Checkpoint. If the Checkpoint, you are using, moved very far away form the Base-Checkpoint, you can contact me, I will train a LoRA against the Checkpoint you are using.

Next step might be adding other LoRAs/Embeddings. The more LoRAs/Embeddings you are using, the less influence on the result has each LoRA, because the SD-engine has to mix the different results of each LoRA/Embedding to the final result.

Use only the LoRAs/Emeddings you really need to produce your result!

Next step might be adding further words to the prompt. The more words are in your prompt, the less is the influence of each word to the result, because the SD-engine has to mix the influence of each word to the final result.

Do not use phrases like 'Award winning' or 'Masterpiece', they have usually no positive influence. If you do not use such words, you do not have to be afraid of ugly images, because usually SD does not produce ugly person. SD produce a statistic mix of all valid training images, and the statistic mix of all faces is usually very nice! All ugly elements are statistically balanced. And there were probably no images in the trainings dataset with phrases like 'Award-winning' or 'Masterpiece'. If you use such phrases, you reduce the influence of the really important words in your prompt.

Keep your prompt always as short as possible!

Further hint: If you want to produce nude girls, think of, that they do not wear skirts or dresses, which can show the centrifugal force of the turn, Give those girls long hair or ponytail, because hair or ponytail can also show centrifugal force.