At first, when I trained many character LoRA and used them with this model, the characters looked almost identical. I thought it was just a feature of the model. Later, I realized that this model is actually very sensitive to the styles labeled in the training set (for example, “1970s style”).

So, if you want to reproduce the original character’s appearance, you need to include the corresponding style from the training set when calling the character. This characteristic doesn’t appear in other models, so it’s hard to call it a drawback. In fact, with other models, adding a style tag changes not only the character’s appearance but also the coloring to match that style. However, this model only restores the character’s facial features without altering the coloring, which is exactly what I want.

In short, if you want modern coloring while still restoring the character’s original look, I recommend using this model (but don’t forget to include the corresponding style from the training set when calling the character).

As an additional note, Anima’s backgrounds and predictions are generally quite mediocre. Although this model has improved the backgrounds, you can still see traces of that mediocrity. A common workaround is to add “photo background,” but in Anima this effect becomes too strong, making the background overly realistic. However, in this model, that same term works perfectly, striking just the right balance.

whats your opinion of Anima released? does it train nicely?

On Forge Neo here, I'm generating getting black boxes when the same same settings (prompts, text encoder and vae) work perfectly fine for Anima base 1.0. Am I missing something?

Edit: A NaN error in Latent, apparently...? Still, how do I fix this?

Thank you my lord. I love this.