Omega pendant = 'horseshoe pendant'. This is a trick to get similar shape. I tried using different tokens for it, but 'horseshoe pendant' is the best I get. The same with skulls on dressbows. Actually they are masks, and I tried tagging them in datasets like skull masks. The results were terrible; they appeared on Alice’s face like a print on a dress, everywhere.
Mostly outfits have pretty bad belt boots (exception London with shoes and Siren with barefeet). If you don't want to struggle with them, I recommend not using keywords for it, and not using full body shot angles. Specifically in realistic or semi-realistic checkpoints/loras. Astrology signs on aprons almost always aren't right shape, I haven't managed to make them really stable yet and probably i won't. If you want to get them, you need manually to draw them.
Adetailer/onlymasked inpaint required.
Special thanks to @CivChan and @SulfurSniffer for tips (advice) and @RedPinkRetro for supporting me many times.
Siren v1.0/Siren v1.1: Use strength 1.0-0.8. The pendant works pretty well in this version too, exept realistic lora/checkpoints in fullbody shot, that will requare inpainting. Anglerfish's skull doesn't really work at all. It seems 'skull' on Pony and SDXL means 'human skull' and they don't know anglerfish, at least I hadn't success with it. As more or less alternative you may use 'anglerfish's skull tied at the back of her waist' with strength around 0.6, it makes the skull bit less human. Illium sometimes appears, sometimes not; in general, don't push much expectation or hope about this part of the outfit. 'Peals' may help to stabilize little round patterns on the dress. In the same time, it will make it more textured. In my prompting for v1.0 on examples, I did make a mistake. I used
'AMRDD' instead of 'AMRSD'. However, it shouldn't have a significant impact on the results.Silk Maiden v1.0/Silk Maiden v1.1: Use strength 1.0-0.8. The pendant works pretty well in this version, mask on her back isn't so, but manageable. I didn't notice any special problems, so just use keywords for things you want to get.
London v.1.1: Use strength 1.0-0.8.
London v1.0: Not much to say about it, this is the simplest version. Use strength around 0.9. The front view often doesn't work well, the apron appears on the back side of the skirt.
Hysteria mode v1.0: In realistic checkpoints/Lora styles, put strength slightly lower than 1.0 to decrease the effect of oversaturation. 'Colored white skin' keep at least 1.3 and 'red sclera' if you want to get at least 1.1. For foreground images, put 'backless dress' in negative. It should also have belts on the back, but it seems the source of the dataset is not sufficient for that. Additionally, don't use it with 'white choker'; it conflicts when trying to make a backless dress and put the pendant on the back side of the neck. With full-body images, the pendant works much worse. Use additional editing. I used Inpaint Sketch and/or 'Inpaint' 'Inpaint only masked' with only the Lora in prompt.
Royal v1.0/Royal v1.1: 'Horseshoe pendant' works a bit worse than in the Classical outfit, because it doesn't have such strong contrast. It often requires additional editing, especially in realistic images. The effect of oversaturation is present, more so than in v1.1, but less than in v1.0, I would guess. With lover strenght the effect should be less. For the side view, use 'skull on dressbow' with a lower weight than 1.0, and 'brown dressbow'.
Classical v1.2: 'Horseshoe pendant' is turning back as keywords. For backside shots use 'backless dress:1.2' in negative. Skull appears very different with different lora and checkpoints, but basically I recommend using 'skull on dressbow' with strength lower than 1.0.
Classical v1.1: In general the same as previous version. Use strength 1.0-0.8, 'necklace' instead of 'horseshoe pendant' and 'blue dress' instead of 'dress'. Skull on bow from her back works a bit worse, but I don't think many people who need this view. Anyway use 'white dressbow, skull' for it, need to manage the strength of them for your case.
Classical v1.0: Use strength 1.0-0.8. The necklace|horseshoe pendant in full-body images sometimes works well, sometimes not so well, but can be fixed with inpainting 'only masked' with a prompt 'horseshoe pendant' and the lora. In half-body images, there usually aren't such problems. The knife|vorpal blade isn't always generated well, but often is fine. Skull on her back should works well with 'skull', you may to use it with lower strength.
I notice the lora have predisposition to oversaturation sometimes. I will try to fix or decrease that effect at least.I think i fixed it in Classical v1.1
Description
Pretty much changes in datasets since v.1.0. The main thing is remade all faces to look like game's character. Besides, there are many small improvements, making generation a bit better than 1.0 but not dramatically.
FAQ
Comments (5)
I guess both kind of tips took a place 😁 Love the Royal 1.1 previews, can't unsee Mila too now 😂
I love McGee's vision of Alice and the grim, whimsical aesthetics 🖤
There's also the Alice: Asylum Design Bible detailing plans for a further installment of the game series, filled to the brim with concept art and intriguing ideas.
In case you want to explore more outfit designs, there is a good amount of inspiration in there.
Thank you so much for the shoutout! 🙏
I looked at part of that bibelot and see it fully. Yes, some of our outfits look really cool, but almost impossible to make some of them in good quality with just using a few pictures. I have two other outfits done (mostly) that aren't published yet. Like this one, for example (https://civitai.com/images/66671280, Steam), because they have unstable important parts just in front view that makes using it almost senseless.
@DemonAlone Yeah, few bits of concept art are likely not enough and would probably require some IPAdapter magic to try and fill the gaps.
That test pic looks very promising already though! Regardless of inaccuracies, it still would create nice results I think.
Do you train with generated images as well? If so I could imagine using different styles while showing the same features might make concepts more precisely understood.
For example a flat 2D comic art version of the same image without any background would get rid of a lot of the possible distractions while reducing important details like the omega necklace or the alchemical/astrological symbols to a more simplistic outlined form. That way any visual ambiguities like shadows or wrinkles in the cloth etc. could be avoided as possible sources of misunderstanding
@RedPinkRetro , That image has just a huge inpainting, those metallic rings (I forgot the names of them) and belts specially are my personal curse. With Pony, I haven't managed to fix them, so I used Flux for it, and still I see how curved they are. I'm trying not to use generated (synthetic) images for training, as I tried a lot on earlier versions of my Kylie Griffin, where little issues that are almost invisible to humans increase with each iteration and result in end is mess. 2D images as a source are mostly not enough in the end to get realistic images. You can see that on my Ruby and Caroll. They work well on 2D and more or less on semi-realistic, but pretty bad on realistic. Marin is better, but still not perfect. Even after trying to fix it, you can see on my images with realistic images of Kylie Griffin (https://civitai.com/models/707419/kylie-griffin-from-extreme-ghostbusters), pay attention to the hair, it looks unwashed. And this is after countless numbers of inpainting of original sources before training, leaving mostly only the general idea of original images. Erasing background is not a good way too, as my Anasheya style dataset almost doesn't contain that. So, if you try to use it without another Lora, or 'draw', you easily could get a kind of Who Framed Roger Rabbit effect. My apologies for texting too many details about it, maybe I'm doing something wrong.