abbuibuibui/Darkness-Lora

Darkness (KonoSuba) — Illustrious SDXL Character LoRA

中文说明 | English (current)

An unofficial character LoRA for Darkness (Lalatina Dustiness Ford) from KonoSuba: God's Blessing on This Wonderful World!, trained on waiIllustriousSDXL v17 (commonly saved as waiIllustriousSDXL_v170). This repository publishes the seven formal epoch files (E1–E7). Comparative evaluation recommends E6 (darkness_rE6S1017-000006.safetensors) as the default, at strength 0.8. The training images and captions are in abbuibuibui/Darkness-Dataset.

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

Files

Interrupted resume leftovers and *-step* / *-state saves are not published. Only formal E1–E7:

Label	File	Role
E1	darkness-000001.safetensors	Alternate. Early; weaker detail.
E2	darkness-000002.safetensors	Alternate. Early.
E3	darkness-000003.safetensors	Alternate. Round1 white-dress scene often leaked armor.
E4	darkness-000004.safetensors	Alternate. Usable identity.
E5	darkness-000005.safetensors	Alternate. Entered the weight matrix; slightly better white-dress retention at 1.0.
E6	darkness_rE6S1017-000006.safetensors	Recommended default. Best balance of portrait, armor, white dress, sitting, and night street at 0.8.
E7	darkness_rE7S1077.safetensors	Alternate. Final formal epoch; very close to E6; 1.0 more armor leak on white dress.

Download E6 for general use.

E6 SHA-256: 3c1f672ae375ce928d4e569a1079a1ca213a6cee9533dd5b82bf22b9d60b3ec2

Showcase

These images are direct E6 @ 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	Armor 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	darkness_konosuba	Blonde hair, blue eyes, long hair, ahoge, red X hairclips
Outfit	darkness_konosuba_armor_outfit	Paladin armor, orange tabard, white pauldrons, red bow
Outfit	darkness_konosuba_white_dress	White / cream formal dress
Outfit	darkness_konosuba_dark_casual	Dark casual looks from the set
Outfit	darkness_konosuba_black_top_orange_skirt	Black top + orange skirt
Outfit	darkness_konosuba_maid_outfit	Maid outfit stills from the set

Recommended: waiIllustriousSDXL v17 or an Illustrious / NoobAI-family model; LoRA strength 0.8 (0.4 is soft; 1.0 can oversaturate portraits and push white-dress prompts toward armor); Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216.

If ADetailer is enabled on the face, put darkness_konosuba plus blonde hair, blue eyes, and ahoge in the face prompt.

Example:

<lora:darkness_rE6S1017-000006:0.8>,
masterpiece, best quality, very aesthetic, absurdres, anime screencap,
1girl, solo, darkness_konosuba, darkness_konosuba_armor_outfit,
blonde hair, blue eyes, long hair, ahoge,
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,
black hair, brown hair, pink hair, red hair, green eyes, purple eyes, blue hair

Why E6

All seven formal checkpoints were first screened at strength 0.8 with the same recipe and fixed seeds on a front portrait, an armor full-body shot, and a white-dress full-body shot. Early files were weaker; E3 often leaked armor into the white-dress screen. Only E5 / E6 / E7 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
E5	Close; white dress slightly safer at 1.0.	Alternate.
E6	Most consistent identity across portrait, armor, white dress, sitting, and night tests at 0.8.	Recommended default at 0.8.
E7	Nearly the same as E6; more armor leak on white dress at 1.0.	Alternate final epoch.

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 formal epochs @ 0.8

Front portrait	Armor full body	White dress

Round 3: E5 / E6 / E7 × 0.4 / 0.6 / 0.8 / 1.0

Front portrait	Three-quarter view

Armor full body	White dress

Training

• Data: 59 manually selected Darkness stills with matching .txt captions. Captions start with darkness_konosuba, <outfit>, keep_tokens = 2, remaining tags shuffled. Outfit counts: armor 35 / white dress 7 / dark casual 7 / black top+orange skirt 5 / maid 5. Published as abbuibuibui/Darkness-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. Planned ~1239 optimizer steps (59 × 3 × 7); training was resumed across interrupted saves but only formal E1–E7 are released.

Related repositories

• Hugging Face dataset: abbuibuibui/Darkness-Dataset
• ModelScope model: abbuibuibui/Darkness-Lora
• ModelScope dataset: abbuibuibui/Darkness-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 1.0 can oversaturate portraits and push darkness_konosuba_white_dress toward armor (especially E6/E7).
• Without an outfit token the armor outfit leaks (35 of 59 stills). Write other clothes explicitly.
• darkness_konosuba_maid_outfit and darkness_konosuba_black_top_orange_skirt 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.