.jpeg)
Frequently Asked Questions (FAQ)
This is a work-in-progress article that answers some of the questions I get most often.
Updated 01/08/2026
Q: Can you make a model of XYZ?
A: I probably can, but there are a few requirements:
It should be an anime/cartoon character (no realistic or semi-realistic characters).
There should be enough training data available. I usually need between 20 and 400 images.
I generally prioritize requests for characters that don't already have a model on Civitai.
Q: How much does a commission cost?
A: Nothing. I don't charge for commissions.
Q: Okay do you charge licenses fees?
A: No, using my model won't cost you anything extra. However, generating NSFW images usually costs yellow buzz by default, and you'll still need to pay the licensing fee for the base model. :/
Q: Can I sell your models?
A: No. Selling my models is prohibited. I'm a strong supporter of the open-source community, where everyone can freely use and share resources.
Q: But can I sell images I generated with your models?
A: No, for the same reason.
That said, who could realistically prove which model was used to generate an image?
Q: I generated images with your model and then trained a new model on those images. Can I sell that model?
A: Yes, you can.
Q: Do you still have the training data for model XYZ? Would you share it?
A: Yes. Just ask.
Q: Can you make an Anima port of your IL model XYZ?
A: Yes.
Q: Can you make an IL port of your Anima model XYZ?
A: Yes.
Q: Can you update your model XYZ?
A: It depends.
If the model was released before 2026, I can usually improve it because my workflow has evolved significantly. I can also update a model if new source material for the character becomes available.
Otherwise, an update is unlikely to make a noticeable difference and would probably be a poor use of time.
Q: Do you make LoRAs for manga characters?
A: Not at the moment.
I've made a few manga character LoRAs before, but over time I found that they often end up looking too AI-generated. The workflow is also much more tedious and time-consuming than anime screenshots.
However, I've written an article about the process, so if you're interested, you can use it to create your own.
Q: How About colored Mangas?
A: That I can do: No tedious colorizing, only cleaning noise. But keep in mind - I still need enough data for a proper model
Q: Can you share your training settings?
A: Sure. Although you can also inspect the .safetensors metadata yourself—it contains pretty much everything you need.
Q: Do you have Discord?
A: I do, but honestly, I'm a workaholic and don't spend much time chatting there.
Q: Can you give me some tips?
A: Of course! Just ask in the chat or leave a comment below.
Q: You're slow! Can't you make models faster?
A: Sorry! This is still just a hobby for me, and I spend a lot of time preparing my datasets: manually gathering images, cleaning noisy data, upscaling, writing proper (partially manual) captions, and more.
If you know of a workflow that could improve my process, I'd be happy to hear about it. 🙂
Q: How can I help speed up things?
A: The training itself isn't very time-consuming—dataset preparation is. So every step I don't have to do speeds up the process. For example, you could provide the images. If they're already cleaned and properly captioned, I can usually finish the model in under an hour. But keep in mind: The dataset decides whether an output is great or bad.
Q: I requested a character, but you never made it!
A: Sorry! Sometimes I forget to add requests to my list, and they slip my mind.
Feel free to send me a reminder via private message or chat.
Q: You're models suck! Pretty bad.
A: ok.
Q: The model quality doesn't seem very good. The face looks blurry and lacks detail.
A: This is a common issue with SDXL models. Full-body generations often have lower-quality faces, which is one of SDXL's biggest weaknesses.
To be honest, part of the issue may also come from the dataset. I often create models for older anime that are only available in SD quality. To accurately reproduce the outfits, I sometimes have to include lower-quality reference images, which can reduce the overall image quality.
I highly recommend using Hires Fix at a minimum, and ADetailer if possible. Another option is to switch to a different base model. For example, Anima produces noticeably better results even without Hires Fix (at least in my experience).
Q: Can you make models for other base models (e.g. FLUX, SD, Pony, WAN, or NoobAI)?
A: It depends. To be honest, I don't currently plan to create LoRAs for older or outdated models such as SD or Pony.
For newer base models, I'd first take a look at what improvements they offer compared to existing ones and whether they have a large enough user base to justify the effort.
I'm interested in looking into WAN, although I believe my hardware doesn't meet the minimum requirements for video generation.
Q: Okay, what's your favorite base model?
A: I really like the stability and level of detail that Illustrious offers. I've been using it as my main base model for about a year, and I plan to continue supporting it for the foreseeable future.
On the other hand, Anima has excellent prompt responsiveness. This is one area where Illustrious sometimes struggles, especially when you're trying to include a lot of different elements in a single image. You often have to use extremely high tag weights, and beyond a certain point (around 3.0 in my experience), they tend to blend together. To get the desired result, you usually have to increase the weights of the tags you want while suppressing the others.