This LyCORIS allows you to generate images with the same style as the animation studio Ufotable.
Version 3.0 has been trained on approximatively 1200 screenshots from the three Fate/Stay Night Heaven's Feel movies, and gave pretty satisfactory results, but everything looked like it came from the Fate/ franchise, which was the main reason why I created the version 4.0.
Version 4.0 of this LyCORIS has been trained on way more images of multiples works from Ufotable.
Firstly, the same screenshots from Heaven's Feel are present, but I added an extra 600 screenshots from the movies.
Then, I added more than 1800 screenshots from Unlimited Blade Works, more than 1200 screenshots from both seasons of the anime Tales of Zestiria the X, 800 screenshots from Kagetsuki/Touken Ranbu, and hundred of screenshots coming from the last three Tales of games openings, God Eater games openings, and some of Fate/Grand Order PV animations which were made by Ufotable.
Unfortunately, with great improvements come a big downside. The LyCORIS seems to have learn that women should have the weird bangs of Alisha, from Tales of Zestiria.
I already tried to retrain the LyCORIS with a hundred images of Alisha removed from the dataset, but it changed nothing.
To remove these weird bangs, I recommend you to add (crossed bangs, bangs, hair between eyes:2.0) to your negative prompt.
The trigger word is ufotable style, and I recommend a weight of 1 for this LyCORIS.
Description
4x more images for training
LyCORIS version
FAQ
Comments (6)
the sdxl version looks like it uses lora not lycoris. at least the txt in the safetensors says it.
It was trained as a LoCon on Kohya_ss webui, which seems to be treated as a LyCORIS on Civitai :)
Any chance you can do a Lora based on Unlimited Budget Works, instead? I've not finished the movies but these don't look nearly as good as Unlimited Blade Works based on your screenshots (not that they're bad but I would prefer UBW style if possible). I'll try this out anyways sometime so thanks.
The v3.0 on Stable Diffusion 1.5 is way better than the v3.0 SDXL. It should also look a lot more like UBW :)
@ToshiBaguette Just checked and you're right, much better. Any idea why or if you can improve it for SDXL?
@kudon44 first of all, I think that I should rework all of the prompts used in the training data.
My training dataset uses dabooru tags as prompts, which is not really good with base SDXL.
I also think that waiting for a better suited model for anime generations could help. The LoCon was trained on base SDXL which, I personally think, is really not that great for anime style images compared to models based on Stable Diffusion 1.5 like AnythingV5 or Counterfeit-3.0.
Then, maybe try to train a new LoRA or LyCORIS could improve the results.
Lastly, I do think that the SDXL is a little bit overtrained. When you ask to generate a woman, it automatically generate Sakura's hair ribbon, and when you ask to generate a man, it goes full Shirou mode.
So I should try to add more images to my dataset (like UBW screen captures, for exemple), and maybe diminish the number of epochs to train the model.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.