Summary

Please consider to donate or sub on my ko-fi here

(all funds go right back into making more loras)

This is a character lora used to generate the character Bowsette (both animated and live action). It can do NSFW without much issue though the current version wasn't trained on nudity (I will do another run of dataset training on that in next version). It is very versatile and works well with other loras including causvid lora (I recommend 12 steps 1cfg no slg/teacache if using it). It is trained on the t2v 14B model so it should also work as an i2v model lora too.

About this version (2.2 WAN 14B)

Trained on both high and low wan 14 T2V model.

I won't do a detailed write up on this. I had tried 3 times when wan 2.2 first came out and it kept giving bad outputs. What I have learned:

High model is super important for character lora. If there is something the base model isn't trained on then it will not work unless it is trained. The advice I'm given contradicts. IF you are training a model for a character likeness (ie a real person's face) then you should undertrain the high, but like with bowsette you wanna train NEW concepts in then you will need a well trained high.

Both is trained to around 100 epochs, I can't recall the steps but I would guess around 10K steps each? I had trained the low 3 times before I gave up... then I came back after learning about wan training for my last 2 loras and found the issue.

Epoch 50 versus Epoch 100 of the high give different results. The lower trained high could not properly create:

The crown with face and pink top, her hair parted in 3 bangs, and the gem placement on her chest. So you need the high lora to create those baselines for the low lora to come in and pick up the details.

This is trained on the exact same datasets as the 2.1 version, and it was one of my first attempts at a character lora. I think I need to come back to this and consolidate some of the captioning (ie remove caption of the crown, horn and her hair so those automatically come in). This lora also will do all the same as the 2.1 so prompt "live action" and put "animated" in the negative and vice versa if you want anime version.

Some of the example generations are blurry due to some setting with my gurren lagann lora I paired it with for fun. Will try to get more example generations in the future. I just want to get this out there so I can move onto some other work.

USE dpm++_sde cfg 3.5 + 5 shift (maybe 6-8 shift is ok), split steps 11/9 between high and low. See notes on the main page for what to tag in the prompt to activate her properly.

About old 2.1 version

Difference between V1 and V1.1 (both is good):

V1 can switch between animated style easier

V2 is trained on additional NSFW data and tends to default to live action but can be fixed with a few extra words in the prompt (see trigger words section)

Necessary Trigger words: BowsetteLORA, Bowsette,

Recommended Strength: 1.0

(See below for more info on prompting)

Dataset

An even dataset of 51 images of Bowsette only (both live action cosplay photos and fan illustrations). Plus v1.1 has additonal 7 NSFW images and double the steps trained.

Resolution 512x768

Main Trigger Words

Necessary Trigger words: BowsetteLORA, Bowsette,

Optional Trigger words for style: animated, live action

I usually chuck "BowsetteLORA" at the start of the prompt and refer to her as "Bowsette" through the rest of the prompt.

Use "animated" or "anime" to trigger animated style, and use "live action" to get real life versions, though the training data is using costumes from cosplayers, so it usually comes out as a professional cosplay if in live action. It can also do 3D CGI style.

Note for animated style in V1.1 only:

V1.1 might require some extra prompting to get animated style, try at the end of the positive prompt as well:
anime style, highly detailed traditional animation, 2D character, bright lineart, stylized lighting

And for negative prompt add

realistic, photo, photorealistic, live action, skin pores, DSLR

Optional Descriptive Trigger Words:

Usually just "Bowsette" is enough and it will pick up on the common items such as the crown and horns, but you may want to specifically prompt certain aspects of her outfit or appearance if they don't appear or you want them specifically. Or sometimes the color is different (ie white horn vs yellow). Every single aspect of her is captioned so you can modify easily. Look below for helpful words to use for prompting reference. (ie you just say "Her tail" or "Green shell on back visible" to put those in). You can also put her in any outfit which is in the base model or other loras or use these in negatives to make them not appear.

Character Features

Hair: Blonde

Ears: Long, pointed, elf-like

Eyes: Large and bright blue, often heavily lined

Mouth/Teeth: Open smile revealing sharp prominent fangs

Nails: Long, black, pointed

Facial Expression: Mischievous, sly, confident, or crazed

Earrings: Blue teardrop-shaped, circular studs, or large spheres

Crown: Gold with a pink domed top, often featuring red gems, sometimes with a star or heart emblem

Horns: Two large, smooth, upward-curving horns (light tan or white), emerging from the hair

Shell: Green spiked turtle shell on her back, often with white trim and tan or white pointed spikes

Tail: Thick reptilian tail, orange or brown, with regularly spaced light tan or grey spikes

Outfit Elements

Top: Black strapless bustier or bodysuit with a sweetheart neckline, usually vinyl/leather-look, with a central oval gem (commonly blue or green)

Bottom: Options include high-cut leotards, short pleated skirts, flowing sheer skirts, or ruffled layered skirts

Stockings: Black thigh-highs (sheer or opaque), sometimes patterned or with thick top bands

Shoes: Heeled sandals, stilettos, or boots — often with spiked ankle straps

Neckwear: Black choker with silver spikes

Armwear: Black spiked wristbands and spiked armlets; sometimes long, shiny, elbow-length arm gloves

Training Info

Trained locally on a 3090 using Diffusion Pipe.

Default settings except:

LR 2e-5, Repeats 5, transformer dtype float8, save_dtype bfloat16, blocks_to_swap 8

Steps: 1400 (epoch 22) for V1, 2700 steps for V2 (epoch 41)

Attached is all the captions and an example workflow under "training data"

I cropped and resized all images using Birme website. Then I removed all watermarks or text from images using gimp, then I fed the images in batches of 5 to google's gemini 2.5 pro (it's amazing at captioning images). I used seruva19's prompt as a base and used that with gemini to get all prompts done. I wanted the captions to be detailed so that you could have flexibility in changing her outfit, style, design etc. but keep the core basics like the crown spiky bracelets etc. It was shocked how well it captioned, very few corrections to be made once I adjusted the initial prompt. Though it did get too context heavy after around 40 images and had to be reprompted. I did a lot by hand, I think this could be automated, but I don't mind doing it, took around 2-3 hours and was a lot less brain rot than captioning for the penis lora I made before...

Example caption
BowsetteLORA, against a plain, light warm pink background. Live action Bowsette has voluminous blonde hair styled in a high ponytail. She wears a gold crown with a purple domed top and visible pink gem details. Two large, smooth, light tan, upward-curving horns emerge from her hair. Blue spherical earrings are in her ears. A black choker with silver spikes is around her neck. Her attire consists of a black, shiny, strapless bodysuit with a silver trim along the sweetheart neckline. She wears long black gloves that extend past her elbows, with a silver button detail near the top edge and white spiked details on the forearm. She also wears black thigh-high stockings with white spiked bands around the top, and black stiletto heels. A green shell with a white trim and prominent, long, white, pointed spikes is on her back. A thick, plush, yellow tail with white pointed spikes extends from beneath the shell. She is standing with her body angled, one hand raised in a claw-like gesture, looking towards the camera. Full shot.

Big Thanks

As always seruva19 Ghibli and Red Line lora post along with training data have been a constant inspiration and source of knowledge for me.

Banodoco discord for always answering my questions on training

Kijai for his amazing nodes and advise on using them.