The State of AingDiffusion XL

Hi,

First of all, I want to thank all of you who have used AingDiffusion XL. Thank you very much for using it.

Secondly, let me express my feelings on how I am really sad that fewer people seem to be fine-tuning Animagine XL 3.1 compared to Animagine XL 3.0. I should say that I feel more lonely than 2-3 months ago when I felt a lot of people fine-tuning and training LoRAs for Animagine. Now, I feel like everyone I know here is trying to either fine-tune "a certain horse model" or trying to make a new model from scratch to beat Animagine (or I should say, making a Proof of Concept with 5.1M images, lol).

Anyway, let's get back on track.

I am writing this article to inform you of the current state of AingDiffusion XL.

AingDiffusion XL's goal right now is to provide a good base for further fine-tuning (because most trainers I know are saying that Animagine XL 3.1 is worse than 3.0), while still doing a good job of generating images. That's why AingDiffusion XL has faster development cycles than most of the models available here on Civitai.

And of course, faster cycles need more funds. It's getting harder to use less than 24 hours' worth of Runpod credits for a fine-tuned iteration. I still haven't replaced my poor RTX 2060, so I still rent cloud GPUs for everything and take it easy on local generations. I live in a third-world country. Our currency has just undergone a severe value depreciation, and it's getting more expensive to train on the cloud.

On the other side, I want to add and add more images to my fine-tuned dataset, so AingDiffusion XL will perform better and better.

But because people have less and less interest in Animagine-based models, it also kind of hurts me a little bit. It's like I spend my salary and pocket money for nothing.

If you're going to leave a comment, "Fine-tune Pony then, like everyone else." Pony still has two big issues: unnatural scoring tags and overtraining. I might fine-tune Pony if I have enough funds to remove those two issues, but for now, no.

Sorry for the long rant in this article.

Thank you for reading this announcement/rant.