Sign In

DCAU Style for Pony XL (DC Animated, Justice League/Unlimited)

163
1.3k
7.7k
78
Verified:
SafeTensor
Type
LoRA
Stats
697
1,973
Reviews
Published
May 1, 2024
Base Model
Pony
Training
Steps: 1,442
Epochs: 17
Usage Tips
Clip Skip: 2
Strength: 1
Trigger Words
dcaustyle
Hash
AutoV2
2C3F2AB0CB
default creator card background decoration
navelgaze's Avatar
navelgaze

This Lora only works on Pony v6 XL and Pony-based models.

Trained on screenshots from Justice League/Unlimited and other DCAU shows with the streamlined animated series designs.

Trigger word: dcaustyle

  • Strength: 1.0 (or lower if too cartoony)

  • Use Pony's quality tags: "score_9, score_8_up, score_7_up," etc. (In some cases, removing these tags can make the style more accurate at the cost of possibly making it more badly drawn.)

  • Use booru tags. Natural language prompting can work too.

  • (With v2 of the lora, the style seems to be unreliable on the base Pony v6 XL model, but when it works, it can be accurate. It's more consistent when used with anime-based Pony merges and loras, though the style may get slightly anime-like.)

Notes

This is my first try at training a style lora. Let me know how it works. Good for solo images. May not be very coherent with complex multi character images. As always, use detailer or inpainting to fix faces that are small in the frame. This was trained on civitai because I don't have enough vram.

Dataset was made almost entirely from screenshots from fancaps.net of JL/U, upscaled TNBA (after the redesign) and the movies that used the Timmverse art style. I put more emphasis on major female characters, but I included some men, random characters, and scenery. I avoided crowd shots or pictures where characters were overlapping, since that seemed to confuse the AI based on a previous attempt. There were also just 2 to 3 generated nsfw images, which was an experiment I might expand on in the future. Nonetheless, it can do nsfw, since this is Pony after all.

Many of the frames from the cartoon are simple and at times "badly drawn". I wanted it to look authentic to the cartoon, not refined, detailed artwork. However, maybe it was a bad idea to train this on "bad" drawings... (Somebody else made a lora that looks more like Bruce Timm's personal artwork: Retro Cartoons, the Animated Series.)

If you know how this dataset can be improved, please give me some advice. I plan to continue upgrading it.

Version 2

Trained on 212 images, up from 128. Added more pictures of the featured characters I'm trying to train and random characters/costumes for dataset diversity. This version has more varied male characters too, though major male characters were not represented by more than a couple of pictures and will probably not be accurate. Sorry.

Some tags used for major characters were changed. See the updated character section below. Training was slightly longer with tweaked settings. (2 repeats x 17 epochs = 34 looks at each image.)

SD 1.5 version

I made this by request, just with mostly the same dataset as v2. It doesn't work that well. Use it with a 1.5 anime model. You'll need to use hires fix, inpainting, detailer, etc. to improve the images from this one. Tags should be the same as v2, except of course don't use Pony's tags.

Characters

This is meant to be a style lora, but I also tried to make it like a multi-character lora, which maybe was trying too much. But it's an experiment. (Other people are already making individual JLU character loras for Pony, which should work with this.)

Some characters were trained with about 6-12 images each, and can be generated with limited accuracy. Try using the name in bold and some of the booru tags used to train them. (You don't have to use all the tags.)

  • Batgirl: batgirl, light blue eyes, mask, red hair, black lips, two-tone cape, belt, black bodysuit, yellow gloves

  • Barbara Gordon: barbara gordon, light blue eyes, no pupils, red hair, black lips, black shirt, pleated skirt

  • Black Canary: black canary, blue eyes, blonde hair, long hair, sidelocks, medium breasts, cropped jacket, black leotard, pantyhose, cleavage

  • Galatea: galatea, blonde hair, bob cut, swept bangs, aqua eyes, gold belt, white leotard, turtleneck leotard, long sleeves, cleavage cutout, gloves

  • Hawkgirl (Justice League season 1-2, mask doesn't always work): hawkgirl, hgmasked, feathered wings, orange hair, long hair, strapless leotard, belt, tights, red footwear

  • Hawkgirl (JLU outfit with no mask): shayera hol, feathered wings, green eyes, constricted pupils, orange hair, parted bangs, long hair, yellow sports bra, yellow track pants, midriff

  • Huntress: huntress, black hair, long hair, black eyes, mask, cape, black leotard, navel cutout, purple belt, utility belt, elbow gloves, thigh boots, thigh holster

  • Lois Lane: lois lane, black hair, swept bangs, purple eyes, earrings, suit jacket, purple jacket, pleated skirt

  • Supergirl: supergirl, blonde hair, swept bangs, light blue eyes, no pupils, hairband, t-shirt, crop top, midriff, cape, miniskirt, gloves

  • Wonder Woman: wonder woman, black hair, long hair, blue eyes, constricted pupils, red lips, tiara, star earrings, star /(symbol/), strapless leotard, lasso, high heel boots, red footwear

  • Vixen: vixen, very short hair, pixie cut, black hair, orange eyes, dark skin, tooth necklace, pendant, tooth earrings, orange bodysuit, cleavage, plunging neckline, belt, elbow gloves, claws

  • Zatanna: zatanna zatara, black hair, long hair, medium breasts, magician, tuxedo, suit jacket, white bowtie, vest, leotard, pantyhose, sleeve cuffs, top hat, high heels

Characters with fewer images. Might work with low accuracy:

  • Harley Quinn (jester costume; might work because the model already knows her): harley quinn, domino mask, blue eyes, black lips, makeup, jester cap, cuffs, two-tone bodysuit

  • Harleen Quinzel (out of costume): harleen quinzel, short twintails, blonde hair, light blue eyes, no pupils, black lips

  • Katma Tui: katma tui, colored skin, black hair, gray eyes, no pupils, orange skin, medium hair

  • Roulette: roulette, black hair, hair bun, hair stick, aqua eyes, dangle earrings, sunglasses, looking over eyewear, tattoo, cleavage cutout, sleeveless dress, elbow gloves, fingerless gloves, side slit

  • Stargirl: stargirl, blonde hair, long hair, blue eyes, star print, mask, long sleeves, midriff, red belt, bike shorts, gloves, holding staff

  • Volcana: volcana, orange hair, very long hair, spiky hair, red eyes, cleavage cutout, long sleeves, midriff, armband, red gloves

Other characters represented by 2-3 pictures in the dataset: Circe, Giganta, Mercy Graves, Poison Ivy. Likely low accuracy.

Concepts that were trained somewhat and might work:

  • pointy breasts: Those classic Bruce Timm pointy breasts, though the TV show did not often depict them. Doesn't always work.

  • gotham city: Use the tags "gotham city, red sky, dark" for that classic red look. If you use "night", it may blend the red sky with blue.

(If you need the old v1 tags, they're here.)