koeri
Aesthetic finetune of Illustrious XL (composed from 0.1 / 1.0 / 2.0 merges) focused on anime-style generation.
Supports artist-style prompting and common tag-style prompting for characters, clothing, scenes, and general anime aesthetics.
Uses v-prediction and ZTSNR sampling, and was trained on ~10,000 curated SFW/NSFW images with Uniform LogSNR.
LoRA-compatible with Illustrious v0.1, v1.0, and v2.0.
Inference
Tag-style: Model was finetuned with general/character tags, comma-separated.
Guidance: 4–6 is a good starting band.
Steps: 20–30 steps.
Resolution: Follow base SDXL/Illustrious defaults.
Styles: LSNet along with the Kaloscope model can query artist tags from any image.
Quality: PerturbedAttentionDelta + QuantileMatchScaling from custom node pack
Training
Base model: latent-space-dreams/LS_Pingo_XL
Source: Illustrious XL merged variants (0.1 / 1.0 / 2.0)
Sampling: Uniform LogSNR
Schedule: v-prediction & ZTSNR
Dataset: ~10,000 curated SFW/NSFW anime images
Trainer: sdxl-trainer (Modified from Aozora)
Training was performed in three passes;
1. around four epochs.
2. one epoch at a lower learning rate.
3. short refinement pass on a curated subset of ~3500 images.
Acknowledgements
latent-space-dreams/LS_Pingo_XL as the base
Resources
Custom nodes for image processing and guidance
License
Description
The dataset is mostly characters, so subsequent versions will have more backgrounds/objects in it.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.