abbuibuibui/echidna-Dataset

Echidna (Re:Zero) — Illustrious SDXL Character LoRA Training Dataset

中文说明 | English (current)

This is the actual training subset used to train abbuibuibui/echidna-Lora, an unofficial Echidna (Witch of Greed) character LoRA on waiIllustriousSDXL v17. The files here are a copy of the directory named in dataset_quality.toml (image_dir = .../03_captioned/main/black_dress). Nothing was added from the 5 holdout images, and captions were not rewritten for this release.

This is a fan-made derivative, not an official product. Rights to the character and source work remain with their respective owners.

Contents

Item	Count
Training images (images/*.png)	25
Matching captions (images/*.txt)	25
Missing same-name captions	0
Orphan captions	0
Holdout images in this repo	0
num_repeats	7

Filenames are unchanged (for example 5077.png / 5077.txt). Captions are comma-separated Danbooru-style tags with the identity and outfit triggers fixed first:

echidna_rezero, echidna_black_dress, 1girl, solo, portrait, looking at viewer, closed mouth, indoors, anime screencap

• Identity trigger: echidna_rezero
• Outfit token: echidna_black_dress
• Training used shuffle_caption = true and keep_tokens = 2, so the first two tokens stay pinned.
• Tags were manually reviewed per image. This set is a single black-dress look; other outfits are not covered as separate trigger words.
• A 5-image holdout split exists locally for later checks, but those files are not included here.

Preview

These nine images are copies of real training stills, not model outputs. They cover face, three-quarter, profile, full body, arms crossed, indoors, cowboy shot, night, and close-up.

Face	Three-quarter	Profile

Full body	Arms crossed	Indoors

Cowboy shot	Night	Close-up

How it was trained

• Base: waiIllustriousSDXL v17, 1024 buckets, bucket_no_upscale
• Network: LoRA rank 32 / alpha 16, UNet plus both text encoders
• Optimizer: AdamW8bit; UNet LR 1e-4, text-encoder LR 5e-5; cosine, 7 epochs, 1225 steps
• Recommended model from that run: E5 (Echidna_ReZero_Q-000005.safetensors) at strength 0.8

Related repositories

• Model: abbuibuibui/echidna-Lora — E1–E7 weights, bilingual card, and the public epoch × strength evaluation.

License

• Character and source-work rights remain with © Tappei Nagatsuki / White Fox / the relevant production committees.
• Dataset files are released under CreativeML OpenRAIL-M.
• Intended for non-commercial fan LoRA training and research.
• Do not use this dataset for unlawful, deceptive, harassing, or exploitative purposes.