(KREA Lora training is expensive, just setting it for 2k total then release, dont tase me bro)
on a pure white background with borderless negative space on str 0.8 or so should omit the gold background
Same dataset but captioned not tagged:
https://civarchive.com/models/1148532/players-handbook-style-or-i-attack-the-darkness-or-il
Will attempt to create semi realistic player's handbook style pictures with a gold border and elements of "frame breaking" on a gold trim border if you ask for it, and often not.
Captioned instead of tagged. Gold border is super hard coded unfortunately. Working on a v2 sans gold borders. You will get a more anime push if you use tags vs captions. I recommend understanding how captioned models prompt like flux, zit, klein to use to it's highest potential.
Str 0.8 to 1 depending on other prompts in the image.
A vibrant digital concept art piece with dynamic brushwork and saturated colors. Absurdres, masterpiece, score_7,
most basic:
illustration of
Works well:
traditional media, or stylized painterly, etc
action shot, dynamic pose, attacking, magic
Anima likes absurdres, masterpiece, score_7, it may or may not help with heavily stylized loras.
Description
FAQ
Comments (5)
<use tags vs captions. I recommend understanding how captioned models prompt like flux, zit, klein to use to it's highest potential>
Is there a good write-up you can point to?
I've always struggled with Flux prompting (I have lots of Opinions(tm) how bad it is. Fucking insane model parser, thinks it knows my own image better than me). But I assumed ZIT's language is more sane.
OTOH I'm winging it, never saw a good guide.
And I can't be the only one -- everyone(i says) uses LLMs for flux prompting!
I have a gemini gem you can use I made that I can run prompts through. Obviously if you're getting really, really down and dirty it wont comply but you can always learn from it.
https://gemini.google.com/gem/1zZdCajoptuqXBjylt4LnHR0GY3Nrx_Hh?usp=sharing
otherwise you can go somewhere like here
@Sophorium Thanks for sharing. Both seem like useful resources (so far. I'll be sure to complain if I decide I need more).
The gemini gem in particular seems to have been well trained. (Very responsive with a minimum of sass. IME it takes a while until a girl gets to that point).
Krea seems to understand natural language a lot better than Flux, but there's a sort of "green zone" for complexity. It gives kinda boring outputs with too little detail, but it starts to miss things with too much. However, it's astouding just how much complexity it can handle before it starts to miss things.
Anima, on the other hand (not sure which one you're wanting to use here), wants booru tagging. It can understand natural language, but it does its best work with booru tags at least mixed in. For that, there's an autocomplete "node" (it really just modifies text input nodes) that's quite helpful.
https://github.com/newtextdoc1111/ComfyUI-Autocomplete-Plus
@Anistasiata yeah anima is much better with tagging at this point I almost never natural language anymore with it unless it's a specific thing I want it's attention on in a special way