tungsten-logotungsten-logo
fennfoto [PONY]
lynodie
lynodie
PDXL checkpoint
·
Uploaded Sep 23, 2024
·
Used 3.7K times
bZtrDDVVZsrp6bfm
nt2FEYSu8bgg6gQ2
N9ksNHQCcizp2X5G
77b4CofwSPb9LjHr
L5R5kEhPmzbRkGad
4e9m7LaQSvcqUL33
2 showcases were hidden due to NSFW settings
About this version
V4 - NONE OF THE EXAMPLES USE ANY ADDITIONAL RESOURCES. You may see tags in the prompts that look like they use extra resources, they do not. I just copy pasted some prompts as they were to see what happened, with the exception of removing negative embeddings from the original prompts. Now for the updates; Faces - I feel like this was the biggest improvement with this update. Many prompts (not all) I tend to get different faces from seed to seed. It's not always the case depending on the prompt, but certainly more than I've seen in my previous versions/other pony models overall. I've trained several datasets on nothing but faces, all with at least 'photo of a girl, 1girl' in the captions to try and cover as wide of an area as possible. But I've trained over 1500 faces, every ethnicity I could think of, multiple different ways, and merged in them several times. When using face detailers, sometimes higher strengths get rid of that facial diversity. When possible I recommend using a lower strength (like 20%) to keep unique results. LORA compatibility - Fictional Character and Style loras work very well from my testing (but the model knows many characters on it's own as well. more than I expected honestly.) I've also worked to TRY and improve 'real person' lora compatibility. I've had mixed results so I'm curious to see what you all can do.
Related Posts