Home Creators Posts Import Register Favorites Logout

Content

These are the models currently sitting on my PC. After testing, most of them usually get deleted, and only two or three remain — the ones I actually use for requests and, of course, just for fun.

Just a quick note: the images below show each model in its default state.

As I mentioned in another post, every model comes with a huge number of built-in styles. Once you start combining them, the possible variations become practically endless.

waiIllustriousSDXL_v160 — sometimes people specifically request this model in their prompts. Without specialized tags, it tends to produce a fairly 3D-looking style, which is why I don't use it very often.

That said, many of the posts featuring shower scenes (😏) were created with this model.

yiffymix_v64NoobXL — this is currently my default model.

The two images above show the difference quite well: the first one is the model in its default setup, while in the second I tried to reduce its overly 3D appearance. The second version is the one I actually use.

Another thing worth mentioning is that this is by far the fastest model in my collection. Image generation is several times quicker than with the models I'll be showing next.

anima_baseV10 — this is actually the model that inspired me to make this post.

You don't need to dive too deeply into the technical details, but the basic idea is this: there are countless AI models out there, yet many of them are built on top of a smaller number of foundational models. Anima is one of those foundations — a base model that other models are trained from.

According to many users, models based on Anima are currently among the best options for generating anime-style artwork. Of course, there's also the whole Midjourney ecosystem, which can produce incredibly high-quality anime images as well, but unlike local models, you can't just install Midjourney on your own PC. 😅

It's also worth mentioning that these models are noticeably slower than something like yiffymix, which itself is based on NoobAI-XL. In fact, they're much slower.

Of course, generation speed also depends on the hardware you're running. My GPU sits somewhere in the mid-range segment, so your experience may vary. Still, speed becomes very important when you're generating large numbers of images.

As an example, for a typical request I generate around 16 images, keeping only about 4 of them depending on how many turn out well. Sometimes that number goes up to 40–50 images before I'm satisfied with the results.

That's one of the reasons these models aren't ideal for handling regular requests. On the other hand, they are trained on a more recent dataset, which means they can sometimes recognize characters that NoobAI-XL-based models (and other older foundations) simply don't know.

Everything below is based on the Anima model.

Before moving on to the next models, I want to show an example of how LoRAs can be used. For those unfamiliar with them, LoRAs are essentially small, specialized models trained on a specific style, character, or concept.

loraMeMaAni_V1 (I'll simply quote the creator's description here):

"This LoRA was trained on images from Megami Magazine from late 2018 to early 2024. Anima already knows the Megami Magazine tag, but it's riddled with Japanese text. These LoRAs were trained with the text manually removed. Please read the About This Version section for more information."

This is the LoRA I currently use most often when creating images just for my own enjoyment. I really like the flat 2D aesthetic because, to me, it feels the closest to the look of traditional anime. Of course, that's entirely a matter of personal taste.

This is still the same anima_baseV10 model, just used together with this LoRA.

You can see how dramatically it changes the style — it almost feels like you're looking at a completely different model.

Next is waiANIMA_v10. The very first model in this list was waiIllustriousSDXL_v160. This one comes from the same creator, but it’s built on top of Anima, whereas the previous model is based on Illustrious.

You can compare the differences in things like color handling, detail level, and overall rendering style.

In general, the key advantage of Anima-based models is how much better they follow prompts. From my subjective perspective, they also handle backgrounds more effectively. Another interesting point is that they can understand text within prompts — meaning you can sometimes specify exact text to be included in the image, although this works with certain limitations.

animayume_v05 — I don’t have much to say about this one, as well as the next models, since I only downloaded them today and haven’t really tested them yet.

cottonanima_base1

Actually, I do have something to say about this one.

miaomiaoHarem_animaBase has a very distinctive style, and it tends to reshape characters to fit that aesthetic. As you can see, Saber here looks quite different from her “canonical” version — more than in most other models.

Still, this model is fun to experiment with. It can produce some really beautiful results. However, it’s not really suitable for prompt-based or “accurate” generation tasks.

Files

Comments

No comments found for this post.