I Was Reincarnated as the 7th Prince – Character LoRA
A LoRA trained on I Was Reincarnated as the 7th Prince so I Can Take My Time Perfecting My Magical Ability.
This model is designed to reproduce a selection of characters from the series while keeping their recognizable appearance, outfits and overall anime aesthetic.
Included characters
Lloyd de Saloum, Sylpha, Tao, Ren, Grim, Jade, Albert de Saloum, Saria de Saloum, Alieze de Saloum, Anastasia, Eris, Escher, Pazuzu
Description
Comments (12)
I don't mean to be rude and only say this from the point of constructive criticism but your style Lora's leave a lot to be desired.
The quality of the characters and style is inconsistent all round with it leaning more towards the poor side
The output obviously won't be super high definition without upscaling but you do get a certain degree of detail.
The other issue I have with it all being 40 steps. Which makes me think it was either undertrained or overtrained by a big margin though I am not expert
Take this Lora for example. The cover image is supposed to be Lloyd but it's not exactly a good rendition of him
Using anima as the base, most Checkpoints and lora tend to be around the 25-35 range with the sweet spot being 30. Though with the inclusion of turbo (even by the official bounce with turbo 1.1) it's supposed to be a powerful model with easier reqs.
I admire and appreciate your goal of wanting to recreate characters and official styles.
Of course the goal is to improve with each version which can't happen without making ones like these
I only say all this as I'm a huge fan of your work, such as Dragonball checkpoint which is really good though also has room for improvement
Again, I hope you do not get disheartened by this as I only wish the best.
Have a great day
Quite the opposite, I actually appreciate this kind of feedback. I don't get many comments pointing out the specific areas I should improve, so constructive criticism is genuinely useful to me.
I've used Illustrious a lot, but I have much less experience with Anima. I started working with Anima because, when I began creating models for it, there were still relatively few LoRAs available compared to other models, and I thought it could be a good place to start and learn.
One of the difficulties with this particular anime is that some of the characters are minors. I have to be careful with the dataset and presentation because otherwise the LoRA can get struck. Finding enough references of adult versions of those characters that still come directly from the original anime is complicated.
I actually generate every showcase image at 40 steps when using the standard Anima model, without Turbo. My goal with the preview images is to show as faithfully as possible what the LoRA itself is capable of, so I deliberately don't use detail-enhancing or other LoRAs on top of it to make the results look better than what users can realistically expect from the LoRA alone.
So feedback like yours is actually helpful rather than discouraging. I'm still learning Anima, and knowing specifically what people think needs improvement gives me something concrete to work on for future versions. :)
@Fakcup That's great to hear.
For the dataset and presentation, I get what you mean, so I'm assessing you are using the onsite tagger and training correct?
If so I have heard that if you let the auto tagger do for characters like Lloyd it becomes an issue. Hence why if you manually tag it shouldn't be an issue then.
As for the 40 steps, I understand your point of view and appreciate the honesty. However adding another thing, the goal of Loras is to be used, if the required steps are high to get desired results it pulls people away from using it.
I'd love to know how do you go about categorising your dataset
@erasedc39149 Actually, no. I use local AI tools (mainly Grok) to help me with the captions, and I train the LoRAs locally with my own trainer.
As for the 40 steps, I honestly don't know if they're actually necessary. That's simply the default value from the Anima workflow I downloaded, so I've been using it without really questioning it until now. I'll definitely experiment with 25–30 steps and compare the results.
For categorising the dataset, I generally identify each character with their own tags and then describe their appearance, outfit, expression, pose, etc. However, I'm currently reviewing that approach because I may actually be over-captioning some of the permanent character traits. I'm experimenting with letting the character tag itself learn more of their core appearance while mainly captioning what changes between images.
@Fakcup ah I see.
I don't think over captioning is necessarily bad per se. Another creator I've talked to told me that anima has an innate bleed problem which you solve by tagging a lot so it doesn't affect multiple characters.
You should checkout Metalchromerex's guide for making a Lora. As well since Metalchromerex is retiring, he's putting all of his datasets up for free. They are all tagged mind you. Have a go through them, that should help you improve your captioning
@Fakcup Also I think the problem is that you're using Grok to caption your dataset and use it to write prompts which is making it overcomplicate everything.
It makes it go away from the danbooru style of promoting which anima was trained on.
Try putting some images on the civit trainer and see how it captions the datset
@erasedc39149 I'll take a look at it.
The main reason I use Grok is for larger datasets. For example, my DBZ checkpoint has around 3,000 images, and I honestly can't see myself captioning every single one manually, one by one 😅
@Fakcup true. Perhaps you could specify Grok to tag it in a certain way.
@erasedc39149 I already ask Grok to create detailed captions using Danbooru-style tags.
@Fakcup really? In that case I can't seem to figure out what's going wrong. 🤔
The prompts for your images, do you write them yourself or with the help of Grok?
@erasedc39149 I use Grok, but I specifically ask it to use tags while also turning the prompt into a more detailed scene. I’ve found that Anima responds much better to prompts that are structured like an actual scene rather than just a simple list of tags.
@erasedc39149 I think the main thing I should try changing is my maximum image resolution settings...