Training run 28, epoch 000014

Dataset 16

Images

Count: 77 images
Sources:
- Danbooru (search terms: splatoon official_artwork)
- Inkipedia
- Splatoon Base
- Splatoon 3 artbook scans
Scanning methodology
- Flatbed scans at 600 dpi
- Descreening: https://github.com/6o6o/fft-descreen
- Despeckling: Unity (1x) upscaling through R-ESRGAN 4x+ Anime6B
- Dedusting: manual painting in an image editor.
Treatment:
- Most source images were either unmodified, or simply cropped to isolate individuals or mask logos/text.
- Special effort was invested in obtaining additional images of individual Octolings.
  - Some images were obtained from Splatoon 3 artbook scans.
  - Some images were obtained by lassoing individuals from scenes, and repainting obscured areas in an image editor. Quality and resolution were enhanced by 4x upscaling through R-ESRGAN 4x+ Anime6B.
Flip augmentation: enabled

Tagging

Methodology: Tags were initially scraped alongside images, or tagged by WD14 (wd-v1-4-swinv2-tagger-v2). Tags were edited by model author for correctness and consistency.
Caption shuffling: All tags are randomly shuffled. No trigger words were used.

Training

Dataset replicates/epoch: 10
Epochs: 14
Batch size: 2
Checkpoint: sd_xl_base_1.0.safetensors
Network configuration: LyCORIS/LoCon (dim=16, alpha=8, conv_dim=8, conv_alpha=1)
LR: 0.0001 for both U-Net and text encoder. Algorithm was "cosine_with_restarts" with 5% warmup.
Optimizer: AdamW

SplatoonXL