Eureka from Eureka Seven (Psalms Of Planets Eureka Seven) trained for Pony XL
V6 - A really weird version. I removed the alternate hairs as they were causing issues and instead I threw in basically any Manga/Doujin image I could find of Eureka. I then minimal tagged each image (10 tags per image) and then let Civitai train it for 7300 steps (If you remove the batch size it's considerably more steps)
V5 - Tried to improve existing content, added two hair variations, added her underwear which appears in some media and tried adding her wings
V4 - Military outfit and flightsuit added (I do not guarantee this one will work. The additions of these two outfits in multiple tests was having effects on the model as a whole. This Epoch I have posted is the best of a bad bunch and I still prefer V3 more
V3 - General improvements
V2 contains an experimental 3 in one LoRA containing her Eureka Seven, Astral Ocean and Hi Evolution 3 Eureka appearances (((Do not use, this 3-1 idea didn't work)))
V1 contains the original Eureka
Description
Removed alternate hairstyles as they were causing issues. I added images from the manga and from doujins to the dataset, this seems to have somewhat helped with her hairclip and collar
I edited a Koikatsu Eureka model to give it her military outfit, bodysuit and underwear and made images using that to try expand the dataset. I did try a full Koikatsu dataset but the results were abysmal as the 3D style of that game seems to cause major issues so I binned that idea
Given this thing have 7300 steps in it, which removing the batch size of 5, is well into the tens of thousands of steps it might well be overtrained but even in its overtrained state everything seems to be appearing to some extent as it should be which is the first time that's happened in a result
Try using it at 0.9, if not flexible try 0.8 but at 0.8 her sleeves were changing colour and at 1.0 one of the test images almost looked to be a copy of one in the dataset so this LoRA may be right on the edge of overtrained
FAQ
Comments (5)
V6 is a weird version, it's Civitai trained with about 10 tags per image then the whole thing was trained to 52 Epochs giving it a pretty big step count for a prodigy trained LoRA
I've got mixed feelings on this one. The positives are that all concepts are being understood and appearing additionally the inclusion of Manga and Doujin images seems to greatly help with her hairclip and collar.
The negatives are that the low number of tags especially for styles and poses seems to bleed into the trigger and at weight 1 the LoRA is definitely showing signs of overtraining
At the very least this test showed that her outfits do seem possible to train
My next ideas for this are -
1 - Adopt @knxo's style of tagging by constructing outfit triggers as one word white_long_sleeve_dress then manually tagging each image with poses, styles background
2 - Follow what some other creators are doing by training the character as their own LoRA then making outfits their own LoRA's you use with the LoRA. The benefits of this are the trainings being separate don't mess with one another, the downside is you have extra LoRA's you need for each outfit
@Hellstorm901 we have the same feeling about the negatives, while the positives are is that I can able to established some prompts that could be match from the show. I have my issues that when applying the other outfit lora for example from Nochekaiser's Lilith cosplay lora and then I applied it to Eureka, and as a result her outfit might have mixed up with her straps and the buttons that have been shown, I hope this would resolve this for her. I think that your ideas would be a good solution to benefit this lora
@ColabThree2 Mixing LoRA’s are always going to cause issues as the way LoRA’s work is by injecting either new data into the checkpoint being used or replacing already understood concepts
If you generate 1girl on a checkpoint it will be any girl
If you generate 1girl using my LoRA it will 99% of the time be Eureka with the 1% being Talho as she’s in the dataset as she wore Eureka’s outfit in an image
If you generate 1girl using another LoRA with mine the Lora’s will fight one another based on the weight of the LoRA and try to make something which is a cross between both
This is the same for any tags in the dataset which are conflicting, dress, bra, hair colour etc
This is a limitation of AI technology as it doesn’t understand prompts like a human does
As for the straps, this is a training issue I’ve seen before in my Refia and I’m A Spider So What Lora’s, the straps in some images are likely getting pruned into her body, I think there’s images of her in the show wearing her dress minus the undershirt and her wearing the undershirt minus the dress so I’ll probably need to use those to try get the straps to be their own tag
Hey, can you do Lulutia (By The Grace Of Gods) and Su Xiao (Ma La Nu Pei)?
If you can put together a dataset for the two characters and send it to me by PM in some sort of download link or via Discord (My name is Hellstorm901, you can find me by searching for it in the Civitai discord) then I should be able to look into it
The general advice for making the dataset
- Use images which show the character and their outfit, images which show only part of an outfit or character are fine too as the AI can connect X and Y if it knows what the thing looks like when everything is together in one image
- Don't worry about resizing images, I do it myself
- Blurry images are fine but don't make the whole dataset be blurry images or the model will end up blurry
- NSFW images are fine
- Any image format JPG, JPEG, etc is fine, I change the images to PNG anyways
- You don't need hundreds of images, 20-30 is fine, I might be able to train with even less if the images are good enough they show the character well
- Source material such as Screencaps, Manga, LN illustrations works best and you can use Manga images with Anime images, any other image is fine as long as it looks close enough to the character but avoid extreme artstyle changes such as Chibis
- No content which violates Civitai Terms of Service
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.