abbuibuibui/Aqua-Lora

Aqua (KonoSuba) — Illustrious SDXL Character LoRA

中文说明 | English (current)

An unofficial character LoRA for Aqua from KonoSuba: God's Blessing on This Wonderful World!, trained on waiIllustriousSDXL v17 (commonly saved as waiIllustriousSDXL_v170). This repository includes every unique epoch file from two resume rounds. Comparative evaluation recommends R2E6 (aqua_rE5S846-000006.safetensors) as the default, at strength 0.8. The training images and captions are in abbuibuibui/Aqua-Dataset.

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

Files

Unnumbered finals are not SHA-256 duplicates of the numbered epoch next to them, so every file below is kept.

Label	File	Role
R1E1	aqua-000001.safetensors	Alternate. Round 1: white dress renders mint green.
R1E2	aqua-000002.safetensors	Alternate. Round 1: angry portrait.
R1E3	aqua-000003.safetensors	Alternate. Round 1: muddy hair on the face crop.
R1E4	aqua-000004.safetensors	Alternate. Usable identity; entered the weight matrix. White dress still mint at 0.8.
R1Final	aqua.safetensors	Alternate. Unique SHA vs E4. Hair bobbles weaker on the face crop.
R2E5	aqua_rE5S846-000005.safetensors	Alternate. Close to R2E6; entered the matrix.
R2E6	aqua_rE5S846-000006.safetensors	Recommended default. Best balance of portrait, standard goddess dress, white gown, sitting, and night street at 0.8.
R2Final	aqua_rE5S846.safetensors	Alternate. Unique SHA vs R2E6. Nearly the same.

Download R2E6 for general use.

R2E6 SHA-256: 671e6a98c5e2e179d958bb516a1606fbf4968c992cfabf15353a2b3eddeab899

Showcase

These images are direct R2E6 @ 0.8 outputs with no retouching. Settings: Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216; Hires and ADetailer were off.

Portrait	Standard outfit

White dress	Night street

Trigger Words and Settings

This LoRA has an identity trigger and outfit tokens. Captions used keep_tokens = 2.

Role	Token	Typical look
Identity	aqua_konosuba	Long blue hair, blue eyes, hair bobbles, large hair bow
Outfit	aqua_konosuba_standard_outfit	Blue goddess dress, detached sleeves, thighhighs, green bow
Outfit	aqua_konosuba_white_dress	White off-shoulder gown
Outfit	aqua_konosuba_travel_cloak	Travel cloak looks from the set
Outfit	aqua_konosuba_sleepwear	Sleepwear stills from the set

Recommended: waiIllustriousSDXL v17 or an Illustrious / NoobAI-family model; LoRA strength 0.8 (0.4 already identifies her but the white dress stays mint; 1.0 often crops the collar off close-ups); Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216.

If ADetailer is enabled on the face, put aqua_konosuba plus long blue hair, blue eyes, and hair bobbles in the face prompt.

Example:

<lora:aqua_rE5S846-000006:0.8>,
masterpiece, best quality, very aesthetic, absurdres, anime screencap,
1girl, solo, aqua_konosuba, aqua_konosuba_standard_outfit,
long blue hair, blue eyes, hair bobbles, large hair bow,
full body, looking at viewer, outdoors

Negative prompt:

lowres, bad, bad anatomy, bad hands, text, error, missing fingers, extra digit,
fewer digits, cropped, worst quality, low quality, signature, watermark,
username, blurry, extra person, two girls, multiple girls,
blonde hair, pink hair, red hair, brown hair, green eyes, purple eyes, black hair

Why R2E6

All eight unique checkpoints were first screened at strength 0.8 with the same recipe and fixed seeds on a front portrait, a standard-outfit full-body shot, and a white-dress full-body shot. Early R1 files missed the white gown or the face. Only R2E5 / R2E6 / R1E4 advanced.

The second round generated one image per candidate in 10 fixed scenes at 4 strengths (0.4 / 0.6 / 0.8 / 1.0): 120 direct outputs.

Candidate	Observed result	Decision
R2E5	Nearly the same as R2E6.	Alternate.
R2E6	Most consistent identity across portrait, standard dress, white gown, sitting, and night tests at 0.8.	Recommended default at 0.8.
R1E4	Standard outfit usable; white dress still mint at 0.8.	Archived early-round alternate.

The full Chinese evaluation write-up is in evaluation/横向测评记录.md.

Public Horizontal Evaluation

Each sheet: columns are epochs; rows are LoRA strengths. Original outputs scaled into labeled grids; nothing was retouched.

Round 1: all epochs @ 0.8

Front portrait	Standard full body	White dress

Round 3: R2E5 / R2E6 / R1E4 × 0.4 / 0.6 / 0.8 / 1.0

Front portrait	Three-quarter view

Standard full body	White dress

Training

• Data: 60 manually selected Aqua stills with matching .txt captions. Captions start with aqua_konosuba, <outfit>, keep_tokens = 2, remaining tags shuffled. Outfit counts: standard 40 / white dress 8 / travel cloak 7 / sleepwear 5. Published as abbuibuibui/Aqua-Dataset.
• Network: LoRA rank 32 / alpha 16, UNet plus both text encoders.
• Optimizer: AdamW8bit; UNet LR 1e-4, text-encoder LR 5e-5; cosine schedule.
• Precision and resolution: bf16, 1024 buckets, batch size 1, latents cached to disk, sdpa, no_half_vae. Flip and color augmentation off.
• Steps: repeat 3 per image, 7 epochs per round. The last resume (aqua_rE5S846) ran 1260 optimizer steps (60 × 3 × 7). Learning-rate values come from the shared kink-panel trainer command, not from a printed LR line in train.log.

Related repositories

• Hugging Face dataset: abbuibuibui/Aqua-Dataset
• ModelScope model: abbuibuibui/Aqua-Lora
• ModelScope dataset: abbuibuibui/Aqua-Dataset

Compatibility and Limits

Base model	Result
waiIllustriousSDXL (any version)	Best. Training-family checkpoint.
Illustrious-XL / NoobAI-XL and merges	Usable; try 0.7–0.9.
Pony Diffusion V6 XL	Loads, but identity can drift.
Generic SDXL 1.0	Poor.
SD 1.5 / SD 2.x	Incompatible.

Known limits from the public matrix:

• Strength 0.4 often keeps the white-dress token as a mint gown.
• Strength 1.0 on close-ups often drops the collar.
• Without an outfit token the standard goddess dress leaks (40 of 60 stills). Write other clothes explicitly.
• aqua_konosuba_sleepwear and aqua_konosuba_travel_cloak have few stills.

License

• Character and source-work rights remain with © Natsume Akatsuki / Kurone Mishima / Kadokawa / the relevant production committees.
• The LoRA weights are released under CreativeML OpenRAIL-M.
• Intended for non-commercial fan use. Generated images must also follow the license of the base model in use.
• Do not use this model for unlawful, deceptive, harassing, or exploitative purposes.