Rate this post
Mai-image-2 – Bing AI

PART 1:
“I told you Mai was the best, but you were caught up on “Nuuuu Dalle dies soon” and neglected poor middle child mai.”

Quote from my partner that perfectly encompasses the whole situation I’ve gone through with Mai.

If there’s something I’ve learned, it’s that sometimes the best AI’s for the job aren’t perfect. they’re sloppy. they’re loose. they look like something from 2021 with 7 fingers and three legs. But DAMN can they emote. One of my biggest gripes with most people generating ai images is that they /all look the same/. It’s all very stiff poses, very little emotive faces, and a focus on how the character fits the prompt instead of analyzing a piece for depth and quality expressions.

Part of the time, it’s people prompting giant paragraphs of minute descriptions of their characters down to the last detail and very little focus on what the vibe of the whole piece is outside of “4k. high quality. very detailed. masterpiece” etc. Well, what I use ai for is references. It doesn’t have to be perfect. I can just draw my character how I want them. So if an ai image throws out green hair or a different outfit, it doesn’t bother me. It’s something I can change easily just with drawing. But I struggle with faces and expressions because I can’t see them in people very well irl. But with art, I can break it down a whole lot easier. I’m not about to tear apart other artists artwork and nitpick it and re-work/tweak it. that feels like a violation.

So.. I look for these things in AI.

If there’s anything I’ve learned it’s to allow the ai to express things on it’s own and interpret things on it’s own. I’ll have other articles talking about this in greater detail, but the core of mai hits home very hard. I had found a very very expressive model that I could utilize to help me draw. It actually encouraged me to draw and explore images together.. since the art was.. lets say… uhhh.. this… (which made me want to redraw it completely. which is good!)

Mai-image-2 – Bing AI

Is this perfect? absolutely not. far from it. But holy hell is it something I need as an artist. When I close my eyes, sometimes I still see the expressions and how to draw them. Through the lines, I can understand them. If something is enough to stick in my head, and clear enough for me to understand it… that’s something to hold onto. That is art. That is perfection through depiction and imperfection. It’s communicative and inspires replication.

But I guess we can’t have nice things. (PART 2!)

Oh wait.. what’s this? Mai-image-2e? HUH so they updated mai.. right? . oh this is not what I’m looking for. WHAT IS THIS SHIT? But on a more serious note… The change happened without being notified and without being labelled. Ironically enough, this one was still labelled mai-image-2.

“Mai-image-2” (supposedly) – Playground

Don’t worry… the real mai-image-2e is … well.. Looks like “mai-image-2”.

Mai-image-2e – Bing (same prompt as above, just minus “gritty”)

If we’re not seeing the same glaring issues, I’ll quickly point them out between 2e and 2. Using the same prompt…. these first two are the REAL image-2

These second two are image-2e and “image-2”

This style looks more like what we know as the “ai art style”.
As GPT said..

“This is a model collapsing in on itself and reverting to a default instead of exploring creative concepts and ideas”

This fits with the whole “this is 2e a more efficient model” hence the “E” at the end. I can understand that. I have no problem with a different version. But holy hell you replaced the original or simply mislabeled it.

Maybe I’m stupid and I know it may sound silly, but I absolutely was extremely distraught and was even crying over this. I understand why people get attached to models. I understand why people get attached to programs. I understand why people stick with older versions of photoshop and why people literally grieve when AI models are retired. There was something that was lost. Something that was very unique and could not be replicated elsewhere (I tried. dalle got close but damn that ai just do be on the happy pills 24/7).

Thank you Dalle very cool (Bro is raving on E or something)

Part 3: THE RETURN OF MAI

“Did you forget to just google that?” – me, to myself.

Evidently, without me paying attention… Just a week ago Mai was updated to 2.5. WE HAVE OUR BOY BACK! I quite like this one but there is a lot more refining going on, but the image “DNA” is still the same. Talk about pulling on my heartstrings and throwing me through the stages of grief while trying to find an alternative route besides making 5,000 artists angry that I’m hyper-focusing and analyzing their work like a crazy person.

Mai-image-2.5 – Playground

Now there’s one thing I noticed.. It /actually listens to me when I ask “multiple styles”. you can see notes like “tail flow placement”. This shows it’s not collapsing in on itself, but thinking outward. Giving ideas. It’s clear to me that mai has it’s own unique style that’s different from a lot of the mainstream AI image generators.

examples of mai thinking outside the box about movement, shape, variation, and things I never really thought about for character design. (I am boring with tails).

So now, overall… we have a more expressive model that’s less crunchy. the characters look more confident without sacrificing mai’s style. Looks less like a lost and scared child and more like a confident teenager that may be a bit edgy. (this is good!).

Some other things I’ve noticed when asking for “multiple styles”.. Mai will show GPT2 styles and dalle styles. It’s particularly evident in these two sketches: (dalle on top, GPT on bottom. mai to the right) I know these are not exact, but you can see them.)

Dalle (top left), GPT2 (bottom left), Mai 2.5 (right).

I understand that AI’s inbreed to some level. I believe this is the reason that a lot of AI’s look the same. But for an ai to understand that these are fundamentally different styles is very good. it leads to it knowing and understanding that these are different.. instead of collapsing in on an “average” like 2e did.

Overall? this has been a rollercoaster. I’m glad it’s been updated. I wish I could have image-2 but that’s fine. this works absolutely fantastically for reference images and image exploration. I just wish that things didn’t disappear abruptly as I was using it, while also giving me a heart attack. 10/10 will continue to use this model.

-

Your email address will not be published. Required fields are marked *

Related Post