Cartethyia is a playable character from Wuthering Waves. Fleurdelys is her alternate form, also encountered in-game as a weekly boss. Leviathan is the main antagonist of game version 2.0 (also a weekly boss).
※ 英語が難しい場合は、以下のタグの説明を理解するためにAI翻訳ツールを活用してください。AIに質問することもできます。
※ 如果您觉得英语有困难,请使用AI翻译工具来理解下面的标签解释。您可以向AI提问。
※ If English is difficult for you, please use an AI translator to understand the tag explanations below. You can ask the AI questions about them.
v5
This version is based on a larger dataset that includes fan art and improved tags. There is still a lot of room for improvement but I ran out of time so I'm releasing it now. Compared to v4 the quality is similar and it's a bit more difficult to use (both overfit and undertrained). I may have deep-fried the text encoder, so try lowering its weight. Sorry if it doesn't work for you. Not really proud of this one, but I hope someone finds it useful. I'll fix it next year. 😅
I recommend using a style LoRA or a model with a strong base style to override the 3d style. Artist tags may not be enough.
Training details are on the "About this version" sidebar section.
Negative prompt:
lowres, low quality, video artifacts, jpeg artifacts, chromatic aberration, film grain, absurdly detailed composition, sketch, red pupils, pink pupils
Characters:
cartethyia, blue eyes, blonde hair, long hair, forehead tattoo, crown of thorns, teardrop earrings, dress, small breasts, tabard, bracelet, anklet, sandals
fleurdelys, blue eyes, blonde hair, long hair, colored inner hair, aqua hair, single horn, floating earrings, spiked halo, dress, large breasts, armor, blue gloves, anklet
fleurdelys (leviathan), dark persona, corruption, adapted costume, covered eyes, grey hair, long hair, colored inner hair, aqua hair, single horn, head fins, teardrop earrings, spiked halo, pale skin, black veins, dress, large breasts, chest jewel, armor, colored extremities, black hands, anklet
Leviathan is somewhat undertrained. Contrary to Danbooru guidelines, I tagged her as an alternate version of Fleurdelys so the unet can better apply its knowledge of her appearance. Lore-wise they are entirely separate characters. The inside of Leviathan's mouth is blue, which I tagged as "blue mouth". Her tongue is not blue though.
Detailer prompts:
portrait, close-up: This prompt focuses on the face.
close-up, eye focus, two-tone eyes, blue eyes, pink eyes, grey pupils
close-up, forehead, forehead tattoo, crown of thorns: This is for her tacet mark.
close-up, object focus, jewelry, fleur-de-lis: This is for her necklace and the weights on her breast curtains.
close-up, clothes focus, dress
close-up, weapon focus, bloodpact/tyrvine/haultir/etc.
close-up, hand focus, bracelet
close-up, foot focus, feet only, footwear focus, sandals, anklet
close-up, foot focus, feet only, barefoot, anklet
Unfortunately her shoes are back to being undertrained again. Anyway, just use whatever works.
Costumes:
cartethyia (formal), official alternate costume, white earrings, flower earrings, silver choker, necklace, white dress, cocktail dress, frilled dress, white bracelet
cartethyia (spring blossom), official alternate costume, head wreath, white earrings, multiple earrings, neck ribbon, bare shoulders, two-tone dress, pink dress, white dress, thigh strap, pink high heels
cartethyia (seaside rendezvous), official alternate costume, hair ornament, multiple earrings, blue choker, halterneck, bare shoulders, white dress, layered dress, long sleeves, yellow bridal gauntlet, single bridal gauntlet, platform sandals
cartethyia (summer serenade), official alternate costume, sun hat, hat flower, pink hairclip, no earrings, white sundress, see-through vest, shoulder bag, orange flip-flops
cartethyia (idol), official alternate costume, alternate hairstyle, long hair, short ponytail, hair bow, blue bow, blue earrings, silver necklace, idol clothes, white camisole, black corset, white shorts, thigh strap, white boots
cartethyia (yogoods), official alternate costume, crown of thorns, teardrop earrings, white choker, white dress, wrist scrunchie
I just made up names for these. They don't work since there's too little data.
If you want to remove or modify parts of her default costume:
adapted costume: An alternate costume that is based on a character's default costume or another existing official costume.
If you want her to wear something entirely different:
alternate costume: A character wearing a costume other than the one they are typically depicted in, typically fanmade/unofficial. These can range from casual or contemporary attire to adaptations of their costume into different forms.
Accessory removal:
no crown
no necklace
no terminal
These might help if putting the items in the negative prompt doesn't work.
Additional characters:
male rover, yellow eyes, black hair, black jacket, black gloves, hand tattoo, black pants, black boots
female rover, yellow eyes, black hair, black gloves, hand tattoo, thigh boots
iuno, blue eyes, blue hair, twintails, very long hair, laurel crown, greco-roman clothes, leg wrap, foot tattoo
phoebe, blue eyes, blonde hair, very long hair, floppy hat, white dress, black gloves, white pantyhose, thigh tattoo, white boots
Rover works well. I included some images of Iuno and Phoebe to help determine when training was finished, but they're undertrained.
Poorly-trained characters:
lupa
phrolova
augusta
cantarella
galbrena
zani
abby
fenrico
These characters are in the dataset but do not have many images. I also didn't bother tagging them in detail.
Weapons:
defier: White/blue/gold sword with swirling blue energy. Cartethyia's signature weapon.
bloodpact: Gold/blue sword. Rover's signature weapon.
haultir: The silver/brass/blue sword used by Rover and Cartethyia.
tyrvine: The brass/white/black sword used by Rover and Cartethyia.
greatsword: The long water blade used by Fleurdelys.
3swords, weapon formation, floating sword: The magic floating swords used by Fleurdelys.
Undertrained. Probably the only acceptable weapon is the greatsword.
Items:
pangu terminal: The gourd-shaped device on Rover's back.
genesis terminal: The lantern skin for Rover's pangu terminal.
fleur-de-lis: The pendant on Cartethyia's necklace and the weights on her breast curtains.
mini crown: Rover's birthday crown.
puppet, hand puppet: Cartethyia's hand puppet.
Locations:
central hub
mindscape: The area inside Rover's mind. Also doubles as the login screen.
birthday: The scene when you choose a character to wish you a happy birthday.
ragunna city
avinoleum
cathedral of mercury, indoors
mt. firmament, outdoors
septimont
septimont, vale of glory
septimont, capitoline hill
sanguis plateaus, hunter's den
sanguis plateaus, tetragon temple
journeying paradise
realm of chaos
asphodel barrens
murmurstown
fabricatorium of the deep
I mainly used these tags to absorb any unwanted background information. Try them out and see if they do anything useful.
Other tags:
cartethyia (hologram): Her appearance when you call her using the terminal.
thinking pose: Right hand to chin, left arm crossed, right elbow resting on back of left hand.
glowing tattoo: An active (yellow-colored) tacet mark.
corrupted gunshot wound, no necklace: The unique wound on Cartethyia's chest when Galbrena shot her.
red portal: The portal used by the Fractsidus.
purple portal: The portal used by Leviathan.
dark tide: Basically just a black/purple body of water.
Glowing tattoo seems to work. I don't know about the rest.
Overall style and quality:
3d: All in-game screenshots and 3d promotional footage are properly tagged as 3d. Use this in the positive or negative prompt to control the style. Or don't use it.
game cg: 2d art displayed in-game.
anime coloring: TV anime style. Tends to produce oversaturated colors.
video artifacts: Video source. Pre-rendered cutscenes and screenshots from promotional videos have noticeable compression artifacts, bloom, and cinematic color grading.
lowres: Original image area below 1 MP. All video sources are considered lowres regardless of resolution due to their lossy nature.
highres: Original image area above 1.92 MP.
absurdres: Original image area above 7.68 MP.
low quality: Fan art that was not bad enough to exclude from the dataset.
awa: Highest-quality fan art according to my taste. This is a NoobAI-XL tag typically used as "very awa".
absurdly detailed composition: Many artists deliberately add unnecessary and excessive details to obscure or distract from the subject. This tag absorbs semantic clutter.
chromatic aberration: A lens defect commonly abused by artists and video games.
film grain: An effect commonly abused by artists.
v4
This version uses a new dataset composed of cutscene/in-game screenshots and promotional material. It's better than v3 in many ways, but also worse in many ways. It's still difficult to change her outfit or hair style, and her swords/shoes are undertrained. I recommend using an upscaler and an i2i detailer pass. Please use the comment section to let me know what problems you have. My standards are high, so I will continue improving it.
This LoRA is compatible with Noob v-pred/eps and Illustrious checkpoints. No upscaler or detailer was used to enhance the preview images. I recommend using "red pupils, pink pupils" in the negative prompt.
Cartethyia Usage
Main tag is "cartethyia".
Face: "blue eyes, grey pupils"
Head: "forehead tattoo, crown of thorns, teardrop earrings"
Hair: "blonde hair"
Outfit: "dress, tabard, stole"
Arms: "bracelet"
Feet: "strappy heels, anklet"
Fleurdelys Usage
Main tag is "fleurdelys".
Face: "blue eyes, grey pupils"
Head: "single horn, spiked halo, floating earrings"
Hair: "blonde hair, colored inner hair, blue hair"
Outfit: "dress, stole"
Arms: "armor"
Feet: "barefoot, anklet"
Other Concepts
"male rover", "female rover": The main character. Undertrained, depends on checkpoint.
"restrained, barbed wire": She is being restrained/injured by thorns. 😢
v3
Cartethyia Usage
Main tag is cartethyia. She is a petite elf girl. Her outfit is very complicated. Her overall color theme is white, blue, black, and gold.
Face:
multicolored eyes, bright pupils: She has blue eyes with magenta accents and light blue pupils. If you have problems with it generating red or pink pupils then includered pupils, pink pupilsin the negative prompt.tacet mark: The vertical black mark on her forehead.
Head:
crown of thorns: Her white and blue-colored thorn crown that partially covers her tacet mark.multiple earrings, blue earrings: She wears three blue teardrop-shaped earrings on each ear.
Hair:
blonde hair: She has blonde hair.long hair: Her hair ends at her butt. Do not include this tag if the camera can not see her hair go below her shoulders.parted bangs, loosely tucked bangs, hair behind ear: She has long bangs that are parted symmetrically in the center and tucked behind her ears.sidelocks: The hair growing in front of her ears that hangs down to her shoulders.braid: She has three braids only visible from behind or the side. Do not include this tag if the camera can not see her braids.
Outfit:
undersuit: The black dress she wears underneath everything and the white apron/tabard over it. All of her clothing attaches to this in some way, usually with straps. The undersuit has a center opening with a black strip of fabric that has a white vine wrapped around it.stole: The two white straps attached to her shoulders which hang down over her chest. They have brass-colored weights at their ends.detached sleeves: The gradient fabric (white to blue) behind her arms that hangs down to her knees. The sleeves have a gold thorn design on them. They are attached at her shoulders with blue straps.
Arms:
bracelet: She wears a white thorn bracelet on both wrists with a similar design to her crown of thorns.
Feet:
strappy heels: She wears white shoes with gold accents. Her shoes combine elements fromblock heelsandplatform sandals, completely exposing her feet and toes. They attach to a gold ring on her ankles with decorative white straps. I did not tag the gold rings asankletbecause I consider them part of her footwear, and the tag may interfere with the actual anklet she wears. The heels have a gold fleur-de-lis design on the back. There is also a design on the sole, but it is unclear.single anklet: She wears a thorn anklet on her left ankle only. It has the same color and design as her bracelets.toes, toe ring: A single gold toe ring on her big toe secures the front of her shoes to her feet.
Fleurdelys Usage
Main tag is fleurdelys. She is an adult version of Cartethyia. Her outfit is similar, with some important differences.
Face:
multicolored eyes, bright pupils: Her eyes are mostly the same as Cartethyia. Maybe a little darker blue.tacet mark: Her tacet mark is active, so it is glowing gold.single horn: A long and thin unicorn-like horn protrudes from the top of her tacet mark. It is gold-colored and has glowing accents.
Head:
spiked halo: A large, disjointed blue thorn halo floats above her head.multiple earrings, floating earring: She wears five glowing white earrings on each ear. Their shape is like a four-pointed star or a cross. They float in a line just below her ear with no evidence of how they are attached.
Hair:
blonde hair: Her outside hair color is blonde.colored inner hair, aqua hair: Her inside hair is aqua-colored. Do not use if her inside hair is not visible, such as from behind.gradient hair: Her inside hair is more strongly colored closer to her head, turning back to blonde near the tips of her hair.long hair, wavy hair: Her hair length is the same as before, but now it becomes wavy after her shoulders and is generally more spread out.hair over eyes, eyes visible through hair: She has long bangs that hang in front of her eyes.
Outfit:
undersuit: The white apron/tabard has ripped down the center, becoming part of herstole. The center opening has frayed edges. The vine has turned from white to blue.stole: Basically the same, although it now has two layers due to the ripped apron/tabard.high-low skirt: The dark blue decorative fabric attached to the back of her undersuit. It extends past her feet and reminds me of a beautiful fish tail.
Body:
breasts: She has very large breasts, comparable to Changli. Instead of this tag, use specific size tags likelarge breastsorhuge breaststo suit your preferences.
Arms:
single pauldron: Protecting her right upper arm is a silvery metallic pauldron composed of overlapping curved plates.single vambrace: Protecting her right forearm is a vambrace, which is flared at the end. Material and design are same as pauldron.single gauntlet: Protecting her right hand is a gauntlet. Material is same as pauldron. The blue skin of her hand can be seen through the segmented metal plates protecting her fingers.black hands, colored extremities: Dark blue corruption is creeping up both of her hands, dying her skin to around her mid-forearm. The color gradient goes from black to blue. If it generates gloves, try usingglovesin the negative prompt.
Feet:
bare feet: She does not wear shoes.single anklet: Same design as before, except the anklet is dark blue.
Swordplay:
sword, holding sword: She wields a dark grey sword in her right hand. The sword is wreathed in flowing water and incorporates organic design elements like vines and thorns. It leaves an attack trail when she swings it.sword guard stanceslashing, attack trail
Expressions:
light frown, glaring: She looks like she's about to call me scum. 🥰
Changelog
See comment section for news.
v4
New dataset: 6567 images of Cartethyia, 1494 images of Fleurdelys
Dataset cutoff: 2025-06-20, includes part 1 of game version 2.4
Switched to OneTrainer
Trained on Noob eps 1.1, so it's generally compatible with v-pred and Illustrious
v3
Update dataset with content from 2.2 PV
Add alternate form:
fleurdelysImprove dataset: feet
Improve tags
v-pred: Do not train with noise offset
v2
Add v-pred support
Improve dataset: face, eyes, torso
Add missing tag:
bright pupilsAdd concept naked stole (similar to naked scarf)
v1
Alpha quality. Needs testing and improvement.
Description
FAQ
Comments (16)
This might be the best LoRA ive ever used. The quality is legendary, and you did this sexy little elf justice. Thank you and keep up the good work!
Thanks! She's very pretty.
Working with such a large dataset is challenging, but I'll continue making improvements until I'm satisfied.
So I did some tests and custom outfits are possible to use but kinda trippy.
You need to lower weight to 0.8 and raise the weight of the outfit lora to atleast 1.5 (sometimes higher depending on complexity). And clothing colors tends to shift into white but it's fixable with higher weights.
But with more skimpier outfits her original clothes slip in more often or maybe it's just overlap with original dress. One thing in particular - breast curtains (idk how to properly name these white straps) seems to haunt almost everywhere.
Thanks, I especially liked your Mobius and Amiya images. Cartethyia's outfit is very "burned-in". I'm diversifying my dataset to try and fix this.
Danbooru calls them "breast curtains" but I called them a "stole". I think "breast curtains" is more correct though, so I'll use that for v5.
v5 progress update:
To solve the overfitting issue I plan to augment my dataset with part of an existing dataset, like deepghs/danbooru2024, then I will attempt a full finetune and a LoRA extract. I've done a finetune and LoRA extract during my failed attempts at v4, so I'm already familiar with the process.
The other main issue is "prior loss": preventing the model from forgetting things it has already learned. Regularization images should help with this.
I used Hydrus to scrape >1300 images of Cartethyia and Fleurdelys from Danbooru and Gelbooru. This number will fall as I review them because many are duplicates, unaesthetic, or otherwise unsuitable for training. I'd estimate around 10-20% of these scraped images are "poisoned art", containing patterns and motifs intended to interfere with AI training. The images are not tagged as such, therefore I need to carefully inspect each image for subtle patterns, watermarks, etc. This is very time-consuming.
It's been a while. Sorry for the lack of timely news. I got distracted.
I have reviewed the first batch of fanart. The quantity of images fell from 1300 to 889; 32% of the dataset did not pass my subjective quality/aesthetic standards. Of the remaining 889 images:
- 316 (36%) are poisoned (unsuitable for training)
- 91 (10%) are lower-quality but still have value for training
- 11 (1%) are masterpieces
- 378 (43%) require manual cleaning to remove artist names/logos/artifacts, crop, mirror, etc
The amount of "poisoned art" was surprising. I learned a lot, and I think it's worth sharing.
Anti-AI tools like Glaze and Nightshade are no longer in common use. I researched artists who make poisoned art, and it looks like the new anti-AI meta began sometime in early 2024. Here are some things that bother me the most:
- Film grain and chromatic aberration.
- Using a rake brush on hair or clothing to create noticeable sets of (usually 6) parallel lines.
- "Semantic clutter": Adding too many details, characters, objects, etc, especially if they distract from or obscure the subject. In other words, it is a purposeful overload of meaningful visual information. For beach scenes this includes excessive lens flare, light rays, sparkles, splashing water droplets, and seagulls. The Danbooru tags most closely related to this concept are "absurdly_detailed_composition" and "abstract_background". I think the intent is to confuse autotaggers.
Examples of artists who poison their art: nixeu, jin_sumire, scottie_(phantom2), lyylinkong, jvn, soare_lucien, chascoby, hanozuku, kyo_no_neko, jianshu, redum4, sw_haze, m3ru, kutu, nokku_(ren_nokku), saidamoosu, reika_(reikadraws), birkai, skaven_gongcheng_shushi, yunoji_(onigiri_yumi09), ru_zhai, elocca, an1ftt, nero_augustus
Filtering out the art created to deliberately sabotage AI training is a huge chore. Sometimes there are edge cases where I'm not certain of the intent, so I have to check other art made by the artist.
The rake brush in particular is such a common motif that it's appearing when I use certain models. You can even find it in official art for Wuthering Waves. There is no tag for it; I can't put it in the negative prompt. This attack was successful. I really want to know where this motif originated and how it spread.
Eventually I would like to write an article on poisoned art. But my free time is very limited so this is low priority.
In the time since I collected the first batch of fanart (2025-07-07), 609 new images of Cartethyia/Fleurdelys have appeared on Danbooru and Gelbooru. I am currently reviewing them. After that I will need to do some manual editing and fix the tags. Additionally, Cartethyia has made some brief appearances in cutscenes and promotional material since her release, so I will include those in the dataset too.
I've reviewed the second batch of fanart. The number of images in this batch fell from 609 to 362. 41% did not meet my standards. Of the remaining 362 images:
- 145 (40%) are poisoned
- 23 (6%) are lower quality
- 5 (1%) are masterpieces
- 111 (31%) require manual cleanup
This should be enough fanart for now, 790 total images (889-316 + 362-145).
Next step is manual cleanup, 489 total images (378+111). Most will be simple manual edits. For more complex edits, Kontext or Qwen will be useful; it's good at removing text, logos, and objects. After that is retagging. And I still need to process my remaining game footage and promotional material (screenshot, crop, tag).
Progress has been slower than I'd like, which is a bit frustrating. But I'll keep marching forward.
I finished cleaning the fanart! I even added new fanart up to 11/11. Dataset is up to 812 images now. Kontext was very useful for making edits.
There is a currently ongoing in-game event where you can take photos of characters during combat. This will be useful for anyone creating LoRAs.
Next is fixing tags and processing the rest of my game footage. I hope to start training soon!
I finished screenshotting the rest of my game footage (starting from game version 2.4 part 2) and promotional material, resulting in an additional 2483 images. Next step is cropping and tagging.
Cartethyia has several different official alternate costumes: a blue and white beach outfit, two different white formal dresses (only one image for one of them!), a pink and white dress, a white sundress with a vest, and a blue and white idol outfit. I'm hoping the training process can accurately capture them. Leviathan is also included in the dataset.
The new game screenshots have been cropped. Some of them I duplicated to obtain more detail with different crop regions. Many images had to be edited using Kontext (later FLUX.2) to remove subtitles and other unnecessary elements. This was especially true of the Leviathan cutscenes.
After editing I used an autotagger on the images. I used an LLM to remove unnecessary tags and then manually verified the removals.
Next step is manually fixing the tags on both the new screenshots (2723 images) and the new Danbooru images (810 images).
Once tagging is done I can finally start training. I really want to release v5 next week, but that's very optimistic. We'll see how it goes.
galaxy brain
Best girl right here ✋🙂🤚
I spent my last 3k yellow buzz that I saved from the golden age of Civitai 2 years ago, before all their policies plunged this great website into decline.
With a lot of effort I finished my project with this great lora from Cartethyia and probably my last work that I do on this website, I will go to pixai. But before saying goodbye to the author of this great lora, there would have been more God if there was a promt for his puppets, but the rest is extremely superior to other loras, with all these results selected at 107 for manual editing and then use them as an image generation palette for a Bot with artificial intelligence in 2 chatbot websites rolplay, and with this I say goodbye and But before saying goodbye to the author of this great parrot, there would have been more God if there was a promt for his puppets, but the rest is extremely superior to other parrots, with all these results selected at 107 for manual editing and then use them as an image generation palette for a Bot with artificial intelligence in 2 chatbot websites rolplay, And with this I say goodbye 👋🫶
Thank you! I'm glad my LoRA was useful! I really like the images you made. The quality is excellent.
Rest in peace, soldier. You will be remembered 🫡
v5 is released!
I didn't have much luck with a full finetune again. Maybe I was too impatient, or I should use different hyperparameters. LoRA with dropout seems to work, but I don't know why and I'm still not satisfied with the results. I think this could benefit from a second training run with a small dataset to focus on undertrained concepts, like swords and shoes. Maybe even masked training. Also I think training the text encoder may be unnecessary. Regularization is something I still need to try. Generalized offset noise may have messed up the colors.
Large datasets are a huge pain to work with. I should just call this "more data = better quality" experiment a failure and train a LoRA using a couple hundred images like normal. I also really want to move on from the SDXL architecture. It feels very constraining compared to what is available now.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.














