She seems super nice IRL, does some paintings, acts well, and is appearing on more movies now!
Play nice, be nice, and please share reviews/images!
Description
Updated tags
FAQ
Comments (6)
Couple notable issues.
First, this lora can't seem to decide what style is it (anime, render, realistic, etc). Poked around with a few types, and didn't find any that it was particularly suited for. This may be caused by the particular checkpoint you trained it with. Got some stuff that looked decent, but it was in that weird midground between render and realistic.
Second, it learned her clothes as well. It's extremely resistant to custom clothing at your recommended strength of 1. Strength of 0.5 seems to allow for 25-50% correct clothing generation. This can be lessened by tagging the dataset images more thoroughly.
Would love to see a v2 with the issues resolved if you're interested in making one.
@ThatPersonThere I've just done a lot of pictures with this one to verify what you said, which I knew isn't so, but I got them out perfectly. It seems an issue with your prompting? Try to follow my examples
@ThatPersonThere No v2 necessary... https://civitai.com/posts/9767640?returnUrl=%2Fmodels%2F526406%3FmodelVersionId%3D622899
@Malessar
Those prompts are a bit simpler than what I typically use. Generated a few more to verify things, as I hadn't tried some of the recurring themes I noticed in your prompts.
If you use clothing prompts right after the trigger prompt, especially if they're vague outfit prompts, it's fairly reliable. More complex prompts (multiple items, 1-4 prompts per item) don't fare as well, especially if used after character adjusting prompts (which need the earlier prompt order, and sometimes heightened weight, to reliably edit past the character lora. Prompt order does matter). Add in a pose prompt or two (if not using controlnet, which I didn't try with this), a background/scene/prop prompt or two, and you start to have issues with the extra learning leaking in. Most character loras can handle actor prompts and maybe a pose prompt being between the trigger and clothing prompts, and then maybe some lighting/weather and some background prompts afterwards in a separate chunk with no issues.
If you keep it simple, and prioritize clothing in prompt order, it's fine. If you want any substantial edits to her (and want to be more specific than using an alternating step blend or a % step blend), the extra learning will show itself.
Based on your prompts, it's fine for what you do. I tend to micromanage the details of images more, which is where the issues show up.
@ThatPersonThere i micromanage and add multiple loras with no issues at all lol idk what youre doing but ive been able to very consistently use it with 3 or 4 other loras and longer prompts. Never had issues.