So those of you who have read many of my articles about AI Art generation know that my primary platform for most of the life of this site has been Bing CoPilot using DALLE-3. DALLE-3 has its quirks and frustrations, but I have always found its ability to replicate art styles unmatched. And you can get visceral, crisp imagery that its hard to get anywhere else without paying a premium.
Microsoft is “retiring” CoPilot/DALLE-3 in the next few weeks. Even though I’ve switched lately to using OpenAI/GPT 1.5 for my workhorse renderer, I still use DALLE quite a bit for scenery, generating characters, experimenting with descriptions, etc. Particularly because a CoPilot account is free while OpenAI requires a pay account that limits renders-per-month.
So now I’m investigating FREE alteratives to DALLE-3 that I can use to replace it. So let’s experiment with a few:
- Bing CoPilot/MAI-Image 2e
- Bing CoPilot/GPT 40
- Google Gemini
For an upcoming episode of Cass in Arcanum, I’ve been sorting out renders of some criminals she and her compatriots will have to contend with. Here’s a basic description of one:
“impressionist oil painting;victorian noir fantasy;long view;a huge half-ogre thug glares at the viewer while slapping a pipe wrench against his palm;He is big, tall, burly and heavyset, muscular, with gray-green skin, bald, and scarred. Mean. Intimidating. He wears a worn tweed vest over a white shirt with rolled sleeves, dark trousers, and scuffed shoes.”
And here’s what the various contenders do with that:
DALLE-3
Note the variety of looks here – an advantage of DALLE when you are experimenting a disadvantage when you want consistent characters.
Bing CoPilot/MAI-Image 2e
With MAI-Image 2e I seem to get stuck with the style. I tweaked the style parameters above, but all them have a very similar aesthetic. When that’s what I want, great! But none of these look anywhere near big and burly enough to pass as a “half-ogre” to me.
Bing CoPilot/GPT-4o
Here’s another one where it gets what I’m asking for, but despite varying the style a fair bit – “impressionist oil panting”, “dark fantasy painting”, etc. – it just doesn’t change the fundamental look all that much.
Google Gemini
Now here there’s some real possibilities. All of these to me are competitive with DALLE in terms of crispness and style, and I really love the feel of impact in that first one with him slapping the wrench into his hand.
Alright. How about something a little different. We’ll try the elf princess Raven.
“impressionist oil painting, fantasy, a young elf princess looks amused at the viewer. She is age 30, russian, with high arched eyebrows, lithe and athletic,straight shoulder-length black hair, blue eye-shadow. She wears a purple sleeveless victorian leather vest, bare arms, tight black pants, black boots, a diaphonous green waist sash, and leather bracers. She has a raven tattoo on her shoulder.”
DALLE-3
DALLE here is pulling out its typical “every woman is a fashion model unless you really push back” tendency, but there’s still plenty of variety here..
Bing CoPilot/MAI-Image 2e
Its certainly got consistency going for it. But it’s really stuck on that style and I can’t seem to push it. That style does work very well for generating characters Cass and Arcanum. And one of these is actually the source image I currently use for Raven. But it wouldn’t fit in as well with some of the previous runs I have done.
Bing CoPilot/GPT-4o
Definitely works the age better than DALLE . I’ll buy this as 30 over some of what DALLE yields. But it still has a very distinctive – “the same artist did all of these” look.
Google Gemini
Once again I’m really liking the crisp style. And if I try to push it like I would DALLE – with “Larry Elmore style painting” – I get the below. Which is quite a credible attempt at exactly what I asked for.
Alright. Let’s get away from people and really push it in fantasy.
“impressionist oil painting, a vicious red dragon glowers at the viewer. It is huge and menacing, with yellow eyes, large wings, sharp claws and a long neck. smoke and flame seeps from its nostrils..”
DALLE-3
A good variety as usual for DALLE – without even trying to push style descriptors too much.
Bing CoPilot/MAI-Image 2e
About what I expected from MAI.
Bing CoPilot/GPT-4o
Feels like more variety here than I got with Raven, but still a very “all the same artist” feel for me.
Google Gemini
Once again that’s a decent array of variation with still good consistency. Not nearly as wild and woolly as DALLE, but in limited testing it feels like that’s because it listens better and not because it can’t be pushed in other directions.
So that’s certainly not an exhaustive shakeout, but it definitely gives me another place to look for a free source of source images for OpenAI/GPT at minimum.
Leave a Reply