Ported most my LORAs to Anima, make sure to scroll right to find all the variants.

Current status:

✅ Ahri, Akali, Gwen, Janna, Kai'sa, Lux, Morgana, Orianna, Prestige Syndra, Senna, Sona, Taliyah, Xayah, Redeemed Xayah, Zoe
❌Jinx, Miss Fortune, Neeko, Quinn, Seraphine, Syndra (normal)

The core captions are grouped for each character, and then each separate caption group is optional.

The lora are trained on OneTrainer, 100 epochs (10% warmup steps) at 0.0002LR using ADAMW Cosine, on batch size 4. Takes me <5 minutes per LORA, the model is small so that helps the batch size and makes the training incredibly fast.

This model base is way more accurate than SDXL finetunes, but with the current checkpoints, it does struggle a bit for faces, which tend to be too generic. I've been able to fix the issue by using <segment:face> on SwarmUI, but I wanted to keep the images free of any hires detailer.
Do note that as of June 2026, a lot of upscalers wash the colors out on Swarm for Anima images (UltraSharp, Remacri etc..); the SeedVR2 Upscaler extension doesn't seem to have the issue.

I've tried to provide both old-school captioning and sentence captioning.

Some notes:
-A lot of datasets include the familiars, you can try to gen them but I didn't particularly try very hard to get them to gen properly

-Ahri, half the dataset has no tails, so they're optional

-Akali, somehow the most affected by the face issue, might retrain

-Gwen, small dataset, SGGwen scissors are a thing but they don't gen nicely

-Janna, oldest and worst captioned dataset

-Morgana, the pairs of wings are rough to generate, use (X:2), but no guarantee they'll appear

-Sona, the detached sleeves are actually optional if you only want the elbow gloves

-Taliyah, the same dataset on IL generates the transparency of the twin cape much better, this likely means poor prompting on my end

-Xayah & Redeemed Xayah, mostly based on Wild rift, hence the purple pantyhose for base; human feet/bird feet allow you to choose what you want