Currently have two versions for PDXL: a Wolves Version for wolf yaoguai in the game, an independent Guangzhi Version.

Guangzhi Independent Version

An individual and stylized version trained mostly on Guangzhi's images, including manually curated/edited synthetic data.

It provides better attire control and resemblance, easier nsfw, default to digital art style, more stylized expressions and shapes for better or worse. (Diversity of 3d screencaps is reduced, so this is better for 2D or 2.5D styles)

This lora can also be used to apply Guangzhi's haircut and outfits to other species.

Usage

Canon Appearance

gz, wolf, solo, male, anthro, mature, white mane hair, white beard, hair bun, hair tie, red eyes

Canon Outfits

• Top: battle robe, pattern robe

• Bottom: white pants

• Arm: red vambrace

• Waist: tassets, waist accessory, pattern blue loincloth

• Legwear: leg wraps

• Weapon: double headed fire blade, polearm

It should be easier to do half-clothed pics than Wolves version. For example, battle robe, tassets, bottomless and other naughty tags should be able to stripe only his pants now. topless, tassets, pattern blue loincloth, white pants can show off his muscles while maintaining some decency.

Other Differences with Wolves Version

• This version uses wolf instead of wolf yaoguai as species tag.

• There is no need to use chinese clothing for outfits.

• Positive quality tags are optional in most cases.

Compatibility

From my test, it's usable on PDXL merges that has retained most furry knowledge, such as autism mix and indigo-xl. However, they might add non-canon countershading fur to the character, quality tags sometimes do that too. If that's the case, try to emphasize effective tags in the merge, for example`grey face, grey body` as positive and white face, white fur as negative.

Known Issues

• A bit overtrained. If the default style is too strong for other LoRAs, set strength lower, 0.8 or 0.6 etc.

• Weird artifacts around the corners of mouth.

• Strong lighting and warm tint. I think part of the problem comes from training on PDXL generations. Could try lower CFG (I've used as low as 2 for Pony Realism).

About Training

940 steps, batch size = 4, net work dim/alpha = 8, scaled Unet/Text Encoder learning rate = 8e-4/4e-4, cosine schedule, adamw optimizer.

Wolves Version

Show some love to the wolf yaoguais who had been your practice dummy in skill preview videos!

This Lora is trained and tested on base PDXL V6. It helps make anthro wolf imitating the general design of wolf guais in Black Myth: Wukong.

Usage

This lora doesn't perfectly reproduce trained wolf guais. Here are the wolf guais included in the dataset and their trigger word/feature tags to help improve resemblance:

• Wolf Archer: lgs, wolf yaoguai, white beard, elderly, long beard, head tuft. Outfits: green shawl, chinese clothing, bracers. Weapon: bow \(weapon\).

• Wolf Scout: lch, wolf yaoguai, hair bun, beard, white eyebrows. Outfits: chinese clothing, headkerchief, hari sticks, arm wraps, knee pads, shoulder pads, necklace, black clothing.

• Wolf Swornsword: ljk, wolf yaoguai, beard, long ponytail. Outfits: hair tie, chinese clothing, white topwear, arm wraps, blue pants, knee pads.

• Wolf Soldier: lxw, wolf yaoguai, beard, long ponytail. Outfits: black topwear, chinese clothing, leather armor, unconvincing armor, arm wraps, necklace, loincloth, chewing stick, pecs, half clothed. Weapon: sword or blade, shield.

• Wolf Stalwart: lls, wolf yaoguai, white beard. Outfits: tan capelet, chewing stick, chinese clothing, arm wraps, loincloth. Weapon: poleaxe.

• Wolf Assassin: lck, wolf yaoguai, black beard, facial scar, bald, hairless. Outfits: tooth necklace, armband, leather armor, unconvincing armor, tribal clothing, baldric.

• Guangzhi: gz, wolf yaoguai, mature, white mane hair, white beard, hair bun. Outfits: chinese clothing, battle robe, gauntlets, hair tie. Weapon: flame blade.

• Old Lingxuzi: lxz, wolf yaoguai, elderly, long beard, white beard, wrinkles, ponytail, long eyebrows. Outfits: black and white robe, bead necklace, hair tie, chinese clothing. He only appeared once as a dead body, so not many data on him.

The 3-letter/2-letter tags at the beginning are meant to be the character trigger word, named after their pinyin name.

Other tags prevalent in dataset: male, solo, anthro, fighting pose, digitigrade, open shirt.

Prompt Example

For painting style:

Positive:
score_9, score_8_up, score_7_up, score_6_up, source_furry, <CharacterTriggerWord>, wolf yaoguai, male, solo, anthro, <OtherTags>

Negative:
blurry, low res, text, source_pony, score_5, score_4, messy,

Optionally neg 3d \(artwork\).

For 3d rendering style:

Positive:
<CharacterTriggerWord>, wolf yaoguai, 3d \(artwork\), male, solo, anthro, <OtherTags>

About Style

This lora uses the 3d \(artwork\) tag for in-game rendering style. Neg 3d \(artwork\) tag can push it even more to drawing styles. For some prompt, PDXL score tags and source_furry are often enough to make painting styles.

Known Issue and Bias

• detailed background bias toward 3d style and nature background.

• Too much lighting in some generations.

• East Asian bias on background.

• chinese clothing will add random accessories due to base model bias.

• Those wolf yaoguais just don't know how to properly wear human clothes and can't help showing off their chests.

• It can generate females but the character style will turn western.

• Weapon and clothes coherency is not guaranteed.

About Training

44 images, batch size 2, 1540 steps, 3e-4/1.5e-5 Unet/TEnc learning rate, cosine schedule, adamw optimizer, net dim 8, net alpha 8.