Lora is multi-tagged, so you don't have to use all the trigger words at once. You can create images with splattered paint or just glitch art. The monochromatic tag keeps the main image within a single color (this doesn't affect the color of glitches and paint splatters, but their color can also be controlled via Prompt). And, of course, you can control the strength of each trigger. Overall, this isn't just a Lora style mimicking certain styles and subgenres, but a fairly flexible tool.
strength 0.7-1.5
Instead of results
SDXL Ilustrios is still unrivaled in terms of creating stylistic adapters, flexibility, and ease of inference. Especially when it comes to local generation. It's still my favorite, and other models will have a long way to go to catch up. Some custom checkpoints are on a completely different level compared to the initial tunes or new models when it comes to anime generation. And this model actually reflects that.
zimage (z-anime) I'd call it a candidate to replace lustrios – in some distant future. It has its own vibe, its own charm, but its negative prompt handling is terrible. And the lack of training data is taking its toll. Although I probably should have used a non-distilled model. It should work better with Lors. Maybe I'll tune it someday.
Anima - although I used the preview 3 model for training and generation, not the release version (I don't think this makes much of a difference, judging by the sample images of the release version). The model is new and therefore rather clumsy. It seems to lean toward pop art and retro art. I don't understand the hype around it, to be honest. It's too monotonous and cheap-looking by default, and when using it, I can't shake the feeling that it's significantly behind what's already out there, at least a year or more (barring a miracle). Currently, it's a good competitor and replacement for SD 1.5, but nothing more. There are some indications that it won't surpass that ceiling. However, it's possible that Anima 2 or Anima 3 will be much better. The developers still have the dataset. And labeled data is already 70% of the new model on any architecture.
However, this LoRa, in my opinion, turned out quite good, and overall, the model trains better than I expected. Although it's difficult to judge with only one trained Lora. I still have the training tools on the PC, so I might train something else. Otherwise, I don't see this model as promising in the near future. That's my current opinion.