Sign In

ElixirProject vs AingDiffusion: An Explanation from the Model Creator

ElixirProject vs AingDiffusion: An Explanation from the Model Creator

[This comparison article is valid for AingDiffusion v15 and ElixirProject v1.1. This article might not be valid anymore in the future.]

You might have noticed that I have released a new series of models called ElixirProject. I write this article to help you differentiate between ElixirProject and AingDiffusion.

Method used

  • AingDiffusion is a merge of a bunch of anime models plus my dataset which is trained with the previous version of AingDiffusion.

  • ElixirProject is a fine-tuning of AnyLoRA, and the dataset used is also fully trained on AnyLoRA.

  • AingDiffusion and ElixirProject might share datasets. However, the datasets are still treated differently.

Image generation

  • AingDiffusion generates more stable images at the CFG scale of ~7, while ElixirProject generates more stable images at the CFG scale of ~12. This issue is more visible on generated images with white backgrounds, especially on AingDiffusion.

    Full image resolution: here.

    If you look at AingDiffusion, the image generated with CFG scale 12 is noisy.

  • ElixirProject might generate consistent-looking images between versions, while AingDiffusion will probably not due to the nature of a merged model.

  • Both models are intended to generate anime-like images, mostly trained on woman characters.

  • Both models will generate different images with the same settings and prompt.

Conclusion

  • AingDiffusion and ElixirProject are two different models with two different methods of training.

  • ElixirProject IS NOT an alternative merge of AingDiffusion.

  • Both models will most likely generate two different images with the same settings and prompt.

  • I will still handle these two models' updates.

Thank you for reading. After I wrote this, I realized how useless this article is for you and you probably wouldn't care. So, thank you.

2

Comments