Since my dataset is from Danbooru, you can find all necessary character information there. To create a character prompt, simply search for the character on Danbooru, select an image featuring their default attire, and then copy all available tags directly into your prompt

Thank you for your support. I have send you massage on ko-fi for the access link

I've listed all the characters I trained in the model details. The list only includes characters I personally trained, and does not include characters already present in the Illustrious base model.
For example, characters from Hoyo games like Zenless Zone Zero, Honkai: Star Rail, and Genshin Impact have been updated to include those available up to the current version.

Note: Some of the newest characters from the current version may not be 100% accurate in their original outfits, as there's still limited fanart available when I trained it

Thanks for the detailed review. You can try using ADetailer to fix the eye details in full body generations. It might not be a perfect solution, but it can help improve the eye quality

1. Yes, 18,000 images were used for the first-phase training.

2. My dataset went through preprocessing, I filtered out low-aesthetic images, bad tags, transparent images, and so on.

3. No real-life images were used. The dataset mainly comes from Danbooru.

4. Both, actually. The 18k dataset already had Danbooru tags. For the second training phase, I used auto-tagging.

5. I'm not sure if there's a "typical" method, since every creator has their own approach. But for reference, my method is the same as the one used in Animagine XL, since I was also one of the devs who trained it.

There’s no universal best method, you’ll need to go through trial and error. Full finetuning is expensive, I rent GPUs on RunPod, and that's one of the reasons not many creators do full finetunes. For starters, pick one GPU that you'll consistently use. Set your batch size as high as possible by the GPU that can handle, ideally using 80–90% of the GPU’s VRAM. Learning rate also depends on batch size and whether you're using multi-GPU or not.

As Hysocs mentioned, it really takes a lot of experimenting to dial in the parameters for full finetuning. Don’t be afraid to iterate, it’s part of the process.

Each version used different base, please read "About this version" for more detail

Thanks for your feedback! I’ll include Natasha from HSR in my dataset for the next update to fix the character issue. If you come across any other characters with similar problems, feel free to let me know so I can improve them in future updates

Thank you for liking my model! The issue with v5.0 is that, in some cases, it can generate body doubles, like duplicated heads or cropped heads. It doesn't happen all the time, but it can occur under certain conditions. That’s why I released v5.1 as an alternative for those affected by this problem. If you haven’t encountered it, that’s great! Both versions share the same character knowledge, so feel free to use whichever one works best for you.

it's been there since v2. You can check the showcase image on v2 or v2.1 for an example prompt for Ayase Seiko

Not sure, I don’t have enough resources to finetune a big model like the 20B parameter Qwen-Image

alternative_Universe Probably not, finetuning huge models with 8–20B parameters is super expensive, and I don’t have that kind of money. Maybe if someone sponsors or donates me to help with the development and training costs