#AI knowledge and Prompts

1 messages · Page 1 of 1 (latest)

west grove
#

Looking at Xanthirra work, Kara and a few others....there are definitely a lot of good knowledge this community has built up. The thing is I hate asking for advice in someone's art post as the discussion will start to take away from the art being shown.

So I am going to try here as a way to collect tips and suggestions and How Tos so people like myself that finds it fun can go to find information. I know people like @wicked lion posted things in her channel as well as others. Please please link them. If you have a suggestion to add please do. The hopes is to have a place to gather this info so everyone can improve. I dont think I seen a central place to just discuss and help others which is where I hope to see this goes.

pliant swallow
#

so i tend to start by just feeding it the in-game description. the systems are pretty good at vibing something on the first try
once i have an image i like, i then hit it with negative prompts to clean it up. and finally i ask for different versions like chibi, or photorealistic, etc

#

this works great for "simple" character portraits

wicked lion
#

Oh well thank you @west grove ! Happy to share, absolutely. My knowledge still holds up I think but I feel like the ChatGPT model keeps being shifted a bit internally. I have some criticisms about what the model does now but would never want that to come across as critiquing anyone's actual artwork. So part of my flow lately has just been trying to fight that and work around what I feel are some default behaviors. (Faces all look similar, sort of messy tiling/artifacting, and loss of quality when reusing/respinning assets).

#

I also tend to drift a bit on styles and go with whatever strikes me at the time 🙂 The one thing that's been consistent is that I try to maintain a library of "composite" headshots for people I render.

#

Even if it's something minor like a hairstyle change, I'll generate a new composite for that.

pliant swallow
#

it has definitely changed internally over the last year. i started hitting it with "pre-prompts" when i feel like it's acting too default. not sure what else to call it, but basically giving it info/context before i give it the prompt/description i need it to draw

wicked lion
#

What I have found helpful lately has been to put this at the start: Generate a PROMPT to render this request: and then it'll create a code blocked prompt that you can fine-tune and preprocess before actually running the render.

#

Also ChatGPT has user-context memory (which does NOT show up under "Memories" in Data Settings) so it knows/assumes some defaults based on prior context which can still pollute your work a bit. You either need to explicitly start a thread with some sort of command forcing a fresh context or use an Incognito chat if you're hitting this.

pliant swallow
#

that's basically what i meant, but said more intelligently 😁

wicked lion
#

Like at this point...even the most basic prompt gets me something that is not entirely inaccurate and I have no explicit "Memories" saved:

image.png
pliant swallow
#

i've experienced the same thing. they save more data than they claim
i primarily use chatgpt, but it's fun to hop onto other llms and give it the same prompt to see what happens

snow kraken
#

Same I do basic character descriptions taking out some of the extras they might be wearing I've given directions like - More realistic, grittier....

low veldt
#

So the way I did thigs is more round about. I started with the character description a long ways back and after eventually getting his face how i wanted i cropped it and saved it as a reference image for chat to use to be consistent

pliant swallow
#

i sort of had the opposite experience when these first came out. the llms had a tough time with the idea of a veiled man 😁