Zeno's Ziggurat

RPG characters with AI image creation

I claim no ownership or copyright of these images whatsoever. You may download and use them for whatever purpose you wish.

Home

Copilot Licensing info

About

AI Basics: Copilot Revisited

So I have traditionally changed my character’s “looks” for either special story events or to mark when they’ve “levelled up” a tier. With the Raccoons I’m trying something different – slowly upgrading their look and gear with every installment. Because this can involve a lot of experimenting I’ve been hesitant to do it with my usual OpenAI render, and I’m always starved for credits there. Re-enter Bing Copilot – which launched this site. They’ve retired DALL-E, but their “Stylized Illustration MAI Image 2e” model does a passable job giving me the kind of wild and stylish work DALL-E used to. But their “Vivid and Natural MAI-Image-2.5-Flash” does two things really well

  1. Consistent Character renders
  2. Taking an existing character render as a source, and adding/removing/changing elements of their outfit.

All of my character models for the Raccoons were actually originally done using MAI-Flash. Including the upgrades. I’ve experimented with using Gemini to do editing as well, but I normally just use that for things like “take this in-game screenshot, remove the characters and game overlays, and turn it into a fantasy painting” to generate scenery reference images.

Svetjana:

Here’s an example of the initial MAI renders from sorting out Svetjana’s look. Note the consistency here – even with no source image and playing with the style. All of these are variations on the prompt below. Added the cheek scar later.

“painterly fantasy painting. A teenage swedish warrior girl poses heroically with a longsword. she is age 18, 6’3″, 180 lbs. braided short blonde hair. flashing blue eyes. muscular build with well-defined muscles. Wearing a long scale-mail shirt, a leather belt, red mini skirt, leather greaves, and leather boots.“

And these are the first few installments of updating her look over time – again all using MAI-Flash and just asking it to change her armor. Literally feeding it the first image below and saying things like “Add steel pauldrons, bracers, and shin-greaves”. Or “give her a steel breastplate”.

Note that literally all of the NPCs and quite a few of the monsters in the Purple Wolf series so far have been rendered by Bing-MAI. It still hasn’t caught up to OpenAI in terms of doing the final renders, but it saves me a lot of OpenAI credits generating reference images.

4 responses to “AI Basics: Copilot Revisited”

  1. atcDave Avatar
    atcDave

    It is interesting to see the overall consistency from a fairly simple description. Its funny, now I haven’t even looked at the site in many months. You apparently can add a reference image? So these are mostly early drafts.
    You listed grieves in the first description, but they don’t seem to actually appear until the last two renders? I still get some of the same behavior at OpenArt, where it will add or subtract certain details based on the *overall* input. So its like you had to add a certain level of gear, armor parts, before it would add the grieves.

    But I’m usually fine with points now at OpenArt. I think its much more efficient than it used to be, unless you want a video or something *very* detailed with the AI turned way up.
    I feel like the biggest limitation anymore is my own imagination, or lack thereof.

    Reply
    1. Zeno Avatar
      Zeno

      Yes. It will let you specify a reference image now. I tweaked the description a number of times while I was playing with it. Sometimes added greaves, sometimes didn’t. Pants, skirt… all toying with the basic look. All of those images are slightly different prompts.

      The main advantage for me of Copilot nowadays is that it doesn’t cost me monthly points. So I tend to do my experimenting in Copilot and then let OpenArt do the actual final renders.

      Reply
      1. atcDave Avatar
        atcDave

        That makes sense. I wonder if the multiple AIs actually expands the range of looks you can get, like i remember how plainly SeeDream rendered all faces a lot alike. It had a distinct bias. I do think Marty has an uncommon look, I wonder if some of that is you deriving him from an AI I haven’t been using.

      2. Zeno Avatar
        Zeno

        Marty is the one that it gave me the most trouble with. Primarily because CoPilot still has the same irritating content filters it always did. Winny being thin, gangly, skinny, etc is fine. Yana being tall, muscular, with defined build. Great. Hans being a short, stout dwarf with commensurate proportions – also fine. But adding “overweight”, “pudgy”, “puffy” or any such descriptor to Marty and it would object. Ended up getting there roundabout by describing him as a slightly nerdy alabama frat boy and taking the hit in 2(4) images still hitting the filter. And that’s what I was going for with him. The “built like an actual heavyset but not really overweight young man” look.

Leave a Reply