Just yesterday Zeno_42 challenged me with the idea of revisiting some older characters/teams, with some of the newer renderers available to us. Well that was a fun idea! In two years the quality of renderers available has improved dramatically.
And as it turns out, I’d been playing around with one such team. Some readers may remember Aias and the Aliens from an Icewind Dale run that started in November 2024. I’m thinking of re-using them in a Gold Box game, at some point in the future. As you may recall, this was the most “out there” sort of team I’ve done. And adapting them to a Gold Box game will present some problems, in part due to backdating to 1E rules.
I had initially thought I’d have this post up “right away”, I’d already done a number of renders with the newer GPT Image 2.5. But some of that was about adapting them to a different game, and *none* of it was about recreating IWD in particular. So my tinkering around got more involved, and I went to bed last night still mulling over how to handle some renders I thought were already done.

Anyway, I have now put together the set I wanted. For all six of The Aliens, I will show their original portrait, that portrait redone with GPT 2.5 in the photo realistic style I tend to favor, and then a new render from the description but NOT with the reference image. Sometimes slightly enhanced to feature or change details… to get things just right.
I’ll also include some images by pairs and one of the whole team. Something that used to be impossible!

Aias in DALL-E 3.
Aias redone in GPT Image 2.5
Aias in GPT Image 2.5 without the reference image.
Ila in DALL-E 3.
Ila redone in GPT Image 2.5.
Ila in GPT Image 2.5 without the reference image.
Husband and wife in GPT Image 2.5.
Jinella in DALL-E 3
Jinella redone in GPT Image 2.5.
Jinella in GPT Image 2.5 without the reference image.
Charity in DALL-E 3
Charity redone in GPT Image 2.5.
Charity in GPT Image 2.5 without the reference image.
Jinella and Charity in GPT Image 2.5.
Jill in DALL-E 3
Jill redone in GPT Image 2.5.
Jill in GPT Image 2.5 without the reference image.
Pikwel in DALL-E 3
Pikwel in GPT Image 2.5.
Pikwell in GPT Image 2.5 without reference image.
Pikwel and Jill in GPT Image 2.5.
Aias and the Aliens in GPT Image 2.5.

The last render is the prize. That’s THE big thing we can do now that we could not two years ago. The pairings are more a demonstration of character stability. We can take those portraits and put them into a scene, and actually expect the scene to look the same as the portrait(s)! Bonus, we can put any combination of characters in those scenes. No doubt, more characters is more work, but it is doable with the current renderers.
Stability used to be such a fight. You could learn words and phrases to *mostly* get the same effect. But there was always a random element, and certain results were just difficult to get however you worded it. This team in particular, so many words had to be used in every description just to get the defining wierdness of the character(s) to look right, there was often not enough space left to actually DO anything with them! Now, since you can add reference images, stability is actually a reasonable goal.

I’d also mention, when I do them in a Gold Box game, this team will be different in several details from the IWD team shown here. The later images of Charity show a change I will likely go with. That is, I’m getting rid of the traditional “mage robes” kind of look and switching to something that looks more like an adventuring mage to me; tunic, pants and belt with pouches.

I’ll also add the disclaimer, I get that a lot of viewers don’t like the photo-realistic look as well. Maybe it lacks the dynamic color and drama we expect in “fantasy art”, or maybe its just purely a “style” thing. But I want to see my characters come to life, to see them breath. I like the gritty realism, often contrasted with a fantastical setting. Bringing *my* characters to life is exactly the thing I find deeply fascinating about these renders.

3 responses to “Aias and the Aliens: Revisited”

  1. Zeno Avatar

    Very interesting comparison. Even when GPT is re-rendering DALL-E images it seems to make them much “crisper” – in a way that looks good here but which I may not always want. Then again, I suspect that can be tweaked by how exactly you ask GPT to do the re-rendering.

    The “no-reference” of Aias makes him look like Sean Connery…

    For Illia it actually notably changed her facial expression – both with and without reference.

    Jinella to me looks the most similar across all renders – but that may be because “Drow-charcoal-black” really dominates the facial shadows/contours.

    Charity’s gold color in the “no-reference” really looks a lot more like “gold the metal” to me than the original. Its a really striking look.

    GPT with no-reference seems to have produced faces for all the girls that look more unique/different from each other than the original. They all had very similar facial features in DALLE – sisters at least. They actually look like different people in straight GPT. Greater variety in facial shapes.

    I’d like to see what it can do trying action shots with Pikwell. Is he supposed to have digitigrade legs? The images make it difficult to tell since none of his are full-length and he’s not exercising his legs.

    Everybody has their own tastes when it comes to art. Photo-realistic isn’t my favorite for fantasy, but that’s just personal preference. And partly because for me photo-realistic makes it harder to mix people/real-world concepts with really out-there fantasy stuff. Trying to do a gibbering mouther for my current run. That would be… a challenge… in a photo-realistic style.

    A lot folks really enjoy an anime-inspired look to their RPG art. Also not my bag. I’m not going to pretend that means anything beyond “Different people have different art style preferences”. It would seem really weird to me to go on someone’s personal blog and criticise/complain about their personal art style choices…

    Liked by 1 person

    1. atcDave Avatar
      atcDave

      Yeah I not only redid the description (as best as I could) for GPT Image, I switched from some variation of “Fantasy Art” to “Realistic photo”. That may be part of why they are sharper.
      I thought it was interesting how dropping the original as a reference typically made them more lifelike. More human proportians and shapes. Although in a few cases (especially Ila and Jill) the initial free renders were too extreme. I think it over-reacted to descriptors like “rounded” and “soft”, I wound up with a couple of chubby warrior women. Those terms had been added to the DALL-E renders to push back against its tendency to make women too model-perfect. So I’d say GPT has fixed some of that bias.
      For the facial shapes, I think its about ethnicities. DALL-E rarely paid much attention to them, I think. At least to say I lean heavily on Northern and Eastern European types, and I think DALL-E mostly lumped those all together. Now if I say “French”, it will look quite different from “Swedish” or “Russian”. Before not so much.

      For now, I think I’m happiest with Charity and Jinella. Both are vivid and fun. The idea of being able to do them with character consistency, WOW!

      Yeah I’m not quite sure which way I’ll go with Pickwel when I use him next. In my mind, he would just *look like* a larger Lemur, maybe 3 feet tall/long. Halfling sized, but maybe spending much time on all fours.
      Maybe its beacause of the clothes? Armor even. I do have “humanoid” in the description, which is maybe not appropriate for what I initially wanted. But I find I like the idea of him looking a little more humanoid as he stands up to interact with his human friends, so maybe I’ll make that more “official” if I ever run a PnP game again!
      It will be some time yet before I get to it. I’m sure I’ll be on Brademante for some months yet. Then back to Diomedes and company to finish that first Gold Box series. So, well into 2027 before I have to decide his final look for his next run.

      Fortunately I’ve not had anyone slam my style choices here. But I’ve had a few at the Beamdog site. I think some of them have already decided AI = Bad, and are eager to criticize what I show. The last one I had, phrased a bunch of criticism of the “the AI did this…” variety. Which I actually had a lot of fun with, because in every case he was not criticizing the AI, he was criticizing MY style and design choices. Like he really disliked Brad’s shield. And blamed the AI for it. I don’t know why it struck me as so much fun to reply with a bunch of “no, that was me” answers. It made me laugh, and in the end I thought it all emphasized the human agency to what I was showing.

      Like

  2. Zeno Avatar

    Yeah. I’m very careful with commenting in public forums about my site nowadays. There’s a lot of folks very worked up about AI – and AI Art in particular. For some excellent reasons – its a messy topic. I’m sympathetic to folks who feel like their work has been stolen and is being used for commercial purposes. I’m less sympathetic when folks get aggressive about situations like ours – where we’re not selling squat, not claiming ownership or creative rights to anything, not claiming to be any kind of artists, and not using it for anything except personal enjoyment.

    I made the mistake of posting about Cassandra in the Arcanum reddit group a while back, and the positive responses seemed to got drowned out by the angry ones. And even with folks defending it I ended up feeling like it was distracting from the purpose of the forum. So yeah. Quite happy to do my own thing away from the AI flame wars.

    Like

Leave a comment