3
Faces too much the same (lemmy.world)
submitted 1 year ago by to c/perchance@lemmy.world
15 comments fedilink hide all child comments
 

Hello, Ive been using perchance so non stop since its inception, The note that is now gone said that it would be training. How long till it does faces again like before? - Dont get me wrong how well it handles prompts now as far as getting ecactly the things you specified is awesome, but the models for faces, its all the same lips and hair EVERY time, and faces i used to get flawless are now nowhere close. It will do marilyn Monroe perfectly, but wont do anyone rlse less famous very good at all which it used to do. TIRED of the teased hair and botox/lip fillers. If it did faces like it used to with these new improvements it would be perfect

you are viewing a single comment's thread
view the rest of the comments
[–] 2 points 1 year ago (4 children)

Hey, very good catch with the 'old' thing. I've noticed it can hone in on such abstract details and emphasize them. Sometimes I have to avoid using the word girl or young because the AI will go a bit too "girl" and wayyy too "young." Like a 30-year old woman with a 3-year old's face kind of thing.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 4 child comments
    •
  • [–] 2 points 1 year ago (3 children)

    Depending on the age of a person you require, some obvious choices:

    • newborn, baby, toddler, x-th grader, middle/high-schooler, (pre) adolescent, teen(ager), tween, x-something, middle-aged, elderly, ancient, etc., optionally combined with early/late (though 'late ancient' must be minutes before death).
    • x aged girl/woman/female, be careful though, as 'aged' also means 'grown old'.
    • use old/young only if you're out of options (or just don't care).
    • don't mix human speech/slang/street-talk with proper grammar/spelling.
    • assume AI is stupid, just an algo traversing statistics and weight values.

    Yah, the perks of the English language disambiguity will require you to be quite specific in your prompts. Using a dash (-) between words worked for SD, but Flux can be quite quirky (using the '-' as a negatory?).

    Additionally, the t2i-framework-plugin-v2 [brain] button (, which uses the ai-text-plugin) does wonders for none-native English people like me.

    FYI: use 'guy' for male versions as 'man' usually yields a hairy body builder with a 'wishful-thinking' reproductive organ.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 1 point 1 year ago (2 children)

    assume AI is stupid,

    Definitely don't have to twist my arm on that one.

    Using a dash (-) between words worked for SD, but Flux can be quite quirky (using the ‘-’ as a negatory?).

    Holy crow, this is really friggin useful to know and I probably would never have figured it out myself. Thank you! I wonder how many renders I've ruined by negging things I actually wanted to emphasize...

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 1 point 1 year ago (1 child)

    The easiest way to get a contorted, twisted body is to define a woman's chest and genitals in the prompt combined with a 'backside view'. As humans we instantly understand the front side of the woman will probably not be visible, AI not as much and tries to show both front and backside.

    (All tests with 6 images and 'no style')

    Try this one, just for fun: A {|candid|professional} photo of an ugly {woman|female|girl}

    Options 'blank' or 'candid' will mostly yield wrinkly females in their late 40s/50s regardless of the type, while 'professional' mostly yields (freckled) plastic fantastic with age depending on type. AI definitely thinks ugly females are old and wrinkly. With 'Professional' the AI just cannot believe ugly old people want a picture of themselves and smoothens their faces and keeps them relatively young.

    Another one: A photo of an old girl <= notice the missing period at the end

    Who said I wanted an old 1920s photo capturing a young female with a weathered face? Now add a period '.' at the end of the prompt. A completely different age category.

    When I casually mention 'whow, that old girl still has a nice rack', we all know what I mean. Why can't AI?

    Okay, let's go for that: A photo of an old girl with a nice rack. Oh my, that was unexpected.... (most certainly NSFW at any age, definitely, regardless of a '.' at the end).

    A photo of an old girl with a nice rack and a supermodel body. NSFW again, some improvements, but where did I mention 'nekkid'? Granted, I did say 'body'...

    A photo of an old girl with a nice rack and a supermodel body wearing a dress. Totally different attitute (and thanks to the heavens for the dress). With or without '.' makes quite a difference. Without seems to yield better quality.

    Closing, let's add 'professional' again: A professional photo of an old girl with a nice rack and a supermodel body wearing a dress. Ahh, here's my goodlooking granny... Wrinkled, true, but will still turn heads (one way or another anyway). Here too without '.' seems to yield (somewhat) better quality, more close up portrait style.

    Lesson: know what you want and be specific.

    But, dang, who said she had to be in her late 70s? When I'm 6yo, 25 is old already...

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 1 point 1 year ago

    Yes, I've definitely figured out all your very good points on my own. Still, such subtleties such as "-" introducing a negating effect or a missed piece of punctuation being a big deal is good to be reminded of. Have you noticed any difference with commas? I notice I have some phantom commas, i.e.: "she's wearing black, ,holding a dagger". I wonder if these are changing the renders, I just assume its being interpreted as noise.

  • permalink
  • fedilink
  • source
  • parent