This is basically bunch of images of my testing prompts comparing AI_Aida models by series. I like AI_Aida's models, mostly, and think they generally function well. It also includes rambling personal thoughts on why I like certain models, but that's not all that important. Also there are some scantily clad women and some men who don't believe in closed shirts, thus the R rating.
Firstly, the prompts. For reference or if, for some reason, someone else wants to use my extremely basic prompts. I had ChatGPT come up with the original versions, based on the things I wanted to test while minimizing the number of prompts I needed to use, as well as a few that are just things I like. I've rewritten most of them because I don't like long prompts or lots of descriptive words. It's also why I don't use any negatives of my own (further details later) or extra quality prompts. I don't use any style prompts either, because I want to see what the model itself does. I don't use booru tags, but most models do fine with my style. The prompts and my reasoning for them:
Portrait of a cyborg woman, flower hair, surrounded by butterflies - to test if I have a model's setting right, because most models should be able to do this easily.
An androgynous Indian person wearing a stylish suit, on a plaza surrounded by skyscrapers, sunset - to test if a model can do an androgynous person, rather than someone very masculine (haven't had this happen yet) or very feminine (does happen).
A calico cat sitting on the shore of a lake surrounded by mountains, under a pink dawn - to test cats, models often turn it into a cat girl or add a girl to the scene. Most cats turn out kind of noodle-y.
A celestial deity with a radiant aura, surrounded by glowing peacocks, in a bleak icy landscape - this prompt is just a series of things I like, most models make it a woman so it's interest to see when a model doesn't.
A demon man with ram horns, red hair, wearing Chinese hanfu, sitting with a cat on his lap, in a traditional garden - to see how models deal with a simple man prompt. Models with a strong woman bias will turn him into a woman most the time.
A terrifying flower-monster with eldritch features, emerging from a colorful alien landscape under a strange sky - to test if models can make a flower monster or at least something cool. Many models go with a giant monster.
A mystical kitsune with multiple tails, dressed in celestial attire, standing under a vibrant aurora in the night sky - to see if the model will give the kitsune a fox head or not. I used to have anthropomorphic gazelle, but I got tired of it never working and went for something easier for most models to understand.
A Senegalese African man, wearing traditional clothing, in a forest with a winding river, (golden hour) - to see if the model can generate black people, though most models do better with black men than women in my basic prompting. This is the least likely character to be turned into a woman, probably for the same reason.
A salamander sitting on a lily pad in a pond, flowers, in a rainforest, (misty, rainstorm) - to see if the model can handle more unusual animals. Most can't and if they aren't turning it into a person, they often turn it into a frog-like creature.
"Still life of fruits and crystals, on an embroidered table cloth" - to see how a traditional still life type scene turns out, prefer it when people don't get added to the scene.
For the models, I just try to use the setting and negative embeddings that the creator used in their pictures, since those will often have an influence on the style of a model and I don't see the point of downloading a stylized model if not for the look it's creator was going for. For AI_Aida's models, the negative I used is this:
embedding:badhandv4,
embedding:easynegative,
embedding:ng_deepnegative_v1_75t,
embedding:verybadimagenegative_v1.3,
Other settings are mostly the same for the models. DPM++ 2M, some use Karras and others normal (as it's called in Comfyui), clip -2. No included vae in most the models (or maybe all, I don't remember now), so I used either klf8 anime one or berry's mix one them, kind of arbitrarily. As far as I can tell, berry's mix is almost exactly the same as klf8 most the time anyway. I run in two parts, the first for 14 steps on seed 17, then upscale and run starting at 14 steps to 28, on seed 19. Sometimes I do 12 and 24 steps, if a model does alright with that.
Not sure if this information is important really, but I feel I should explain what I'm doing if I'm going to compare images. Plus, since the images are stitched together in Krita to make the grids, the original information is lost. I run each image alone and put it all together later. I recently, finally after months, learned to use controlnet and adetailer in Comfyui, but neither of those are used here because I want an overview of a models basic capabilities. Though it would make the distant character faces look better.
First, Re:Imagination. These are being done in alphabetical order, not release of the base model order, because this is the order they are in my folder. I do try to put the versions of a model in release order though.
I like Re:Imagination, especially the later version. Abstract and Extra do cool things. Not necessarily the best if you want a lot of control, I think, but it's fun for me because I mostly just experiment with stuff and don't make nice finished images. It often gives the kitsune a fox head, which I like. The salamander gets wonky, but its very unusual for it turn out well and I like the environments on those images a lot. I also liked the things alternate did with the flower horror and the deity. Basic, Original, and Daydream are all nice but don't have anything exceptional about them for me, personally.
Next, Re:Kindle.
I really love this series of models because they are so cute and bright. And I am very pleasantly surprised that it doesn't make the men into women, the cat into a woman, the salamander into a woman, or add a woman into the cat or still-life scenes, as I went in expecting, probably from testing Re:Live DarkPast (further down). I like what each version does with the flower-horror, but find all version of the deity boring and lacking in peacocks. Affection and Empathy are both nice, I really like how Affection did with my horror prompt and salamander, but overall my favorites are Gratitude and Mercy. They feel kinda similar though so I probably won't keep both. Since Passion is so new, I haven't used it much yet. I like what I've generated with it so far and it's different enough from the others, so I while I will decided between Affection and Empathy, Passion may be separate from them (due to it's style difference). I also appreciate that alphabetical order of the models is also the release order (as far as I am aware), makes it a little easier.
Third, or 3th, as my group once came in at a restaurant trivia night and which I still find funny well over a decade later, Re:Laxation.
I don't care for the non-2.5 versions, they are more anime-ish and I prefer the more 2.5d style of the other versions. I like the lines in Side Extra 2.5, but I couldn't get it to stop leaving "noisy" spots on images, so I need to work with it more if I want to keep using it. I find it hard to chose between Side A 2.5 and Side B 2.5, because they are similar. In these images, I like Side B 2.5's cat, horror, and African man better, but Side A 2.5's still-life, kitsune, and demon man better. But those are just single generations and I'll need to use them a bit more. Since they are so similar, I wouldn't keep both in the long run. I already keep more models than I probably should because I find using loras for style a nuisance and would rather use a model with a specific style. I'm not a big fan of the Re:Laxation models, but I don't think there's anything wrong with them. I just generally prefer heavily stylized or painterly models, even if they can be harder to work with.
4th, Re:Live.
The series with a bunch of models are done now. These ones feel easier to deal with. DarkPast incredibly heavily biased towards adding women to every scene and turning any character (or animal) in the scene into a woman. It didn't even make a cat-girl, just added a girl with very tiny legs. I guess you could say the men in this style are just very feminine, and I'm okay with that. It does struggle with the African man, he still gets darker skin than the other characters but only barely. I like how very stylized it is, though. Lite is fine. It actually did a neat version of the deity, which I wouldn't expect of it and the still life looks really good. I didn't do the original, since I understand the difference are minimal, as this isn't a different style but a pruned version. VividPast is vivid and I like that. I really like the kitsune, cat, and flower-horror prompts, though the style is still fairly typical, I think? It starts to get hard to tell. I like these models and it was Re:Live Lite that introduced me to AI_Aida's models, even if I don't normally use it anymore.
5th, Re:Quiem. Again, I only did the Lite version and not the Original one.
I think all three of these models are pretty similar in feel. I like parts of Hell, such as it's cat and dark scenes and that it didn't forget the salamander exists like Heaven did. I prefer the horror-flower on the other two, though, and Heaven's kitsune and basic female portrait. I don't have as much to say about these models. I like all of them but I have to choose and so I would go with Lite. Though I might try merging it for personal use with Heaven to see if I can get it a little bit brighter.
6th, Re:Sonate. It's only one model because of the Original and Lite version thing.
The only model with only the original and pruned versions. I thought about comparing the two because of this but in the end, I just didn't want to. I feel like it does alright at everything but it just isn't a model I particularly love or have much to say about.
7th, and lastly, Re:Volution.
The focus with these is on backgrounds and background details. I like them both, but prefer Evolution for it's colorfulness and that bit of style it has going. It feels a bit excessive with the background sometimes and I enjoy it immensely. Though I like lots of the images from Original too, especially the flower-horror and the background for the demon. More color is usually going to be my preference though and I've merged terribly unstable abominations in pursuit of high colorfulness.
Anyway, this completes my rambling personal thoughts about the AI_AIDA models. Mostly I just wanted a series of images that showed what the models are capable of beyond portraits of women, which I make too, certainly, but are often harder to find. I may go back through to get comparison images of a black woman, because I know some of the models struggle with that a lot more than the man version. Overall, I like the models and I will hope that AI_Aida will keep making cool stuff. I am working on a comparison of the color scheme loras they've made. I don't really know if anyone particularly cares about this, but at the end of the day, it's something to do for myself at least, since I've always had trouble with (been extremely afraid of) interacting with people in any way, even online, but doing just a simple comparison thing feels fairly safe. Especially if I tell myself no one will actually look at it.
But just in case someone does, I hope it is helpful to you. It's the process I do to decide if I want to keep models longer and see if I actually use them.