abbuibuibui/echidna-Lora

Echidna (Re:Zero) — Illustrious SDXL Character LoRA

中文说明 | English (current)

An unofficial character LoRA for Echidna (the Witch of Greed) from Re:Zero − Starting Life in Another World, trained on waiIllustriousSDXL v17 (commonly saved as waiIllustriousSDXL_v170). This repository includes complete E1–E7 weights. Comparative evaluation recommends E5 as the default, at strength 0.8. The training images and captions are in abbuibuibui/echidna-Dataset.

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

Files

Epoch	File	Role
E1	Echidna_ReZero_Q-000001.safetensors	Alternate. Screened out in round 1: heavier early style.
E2	Echidna_ReZero_Q-000002.safetensors	Alternate. Screened out in round 1: mid-transition identity.
E3	Echidna_ReZero_Q-000003.safetensors	Alternate. Usable identity; less bloom at high strength than later epochs.
E4	Echidna_ReZero_Q-000004.safetensors	Alternate. Screened out in round 1: full-body dress detail regresses.
E5	Echidna_ReZero_Q-000005.safetensors	Recommended default. Best balance of portrait, profile, black dress, sitting, and night scenes at 0.6–0.8.
E6	Echidna_ReZero_Q-000006.safetensors	Alternate. Not advanced to the weight matrix. Nearly the same as E7.
E7	Echidna_ReZero_Q.safetensors	Alternate. Final-epoch filename has no -000007. Close to E5; 1.0 blooms more easily.

Download E5 for general use. The other six files are kept so other developers can compare epochs or pick a different trade-off. E3 is the most useful alternate when higher strength headroom matters.

E5 SHA-256: F3FC3F086B1EE56740DC11C9D5186C2A795D04C85DA8B8B266F520D436120496

Showcase

Both images are direct E5 @ 0.8 outputs with no retouching. Settings: Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216; Hires and ADetailer were off.

Black dress (three-quarter)	Street scene

Trigger Words and Settings

This LoRA has an identity trigger and an outfit token. The training set is a single black-dress look; writing only echidna_rezero usually brings the black dress back.

Role	Token	Typical look
Identity	echidna_rezero	Long white hair, white eyelashes / eyebrows, dark eyes, green-black butterfly hair ornament, pale skin
Outfit	echidna_black_dress	High-collar black dress, scalloped collar trim, long sleeves, long skirt

Recommended: waiIllustriousSDXL v17 or an Illustrious / NoobAI-family model; LoRA strength 0.8 (0.6 already identifies the character; collar and chest details are more stable at 0.8; 1.0 often adds white-hair bloom); Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216.

If ADetailer is enabled on the face, put echidna_rezero plus white hair, white eyelashes, dark eyes, and the butterfly ornament in the face prompt, or the face can drift.

Example:

<lora:Echidna_ReZero_Q-000005:0.8>,
masterpiece, best quality, very aesthetic, absurdres, anime screencap,
1girl, solo, echidna_rezero, echidna_black_dress,
long white hair, white eyelashes, dark eyes, butterfly hair ornament,
black dress, high collar,
upper body, looking at viewer, indoors

Negative prompt:

lowres, 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

Why E5

E5 was not chosen because it was a later epoch. All seven checkpoints were first screened at strength 0.8 with the same recipe and fixed seeds on a front portrait and a casual-clothes full-body shot (14 images). E1 looked heavy, E2 was still mid-transition, and E4 lost dress detail on the full-body test. Only E3 / E5 / E7 advanced.

The second round generated one image per candidate in 9 fixed scenes at 4 strengths (0.4 / 0.6 / 0.8 / 1.0): 108 direct outputs. The comparison covered white hair, white lashes, the butterfly ornament, high-collar black dress; profile and full-body proportion; whether casual prompts escaped the default dress; sitting and night-scene identity; default-outfit leak without clothing words; and trigger isolation without echidna_rezero.

Candidate	Observed result	Decision
E3	Usable identity; less bloom at high strength; portrait detail slightly behind E5.	Alternate.
E5	Most consistent identity across portrait, profile, black dress, sitting, and night tests at 0.6–0.8; trigger isolation is clean.	Recommended default at 0.8.
E7	Nearly the same as E5, with more bloom at 1.0 and no stable gain.	Archived alternate.

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

Public Horizontal Evaluation

Each sheet below is a real comparison from one scenario: columns are epochs; rows are LoRA strengths. Images are original outputs scaled into labeled grids; nothing was retouched or replaced.

Round 1: E1–E7 @ 0.8

Left to right: E1 through E7. This round only narrowed the candidates.

Front portrait	Casual full body (still often leaks black dress)

Round 2: E3 / E5 / E7 × 0.4 / 0.6 / 0.8 / 1.0

Left to right: E3, E5, E7. Top to bottom: 0.4, 0.6, 0.8, 1.0.

Identity and viewing angle

Front portrait	Three-quarter view

At 0.4 all three candidates look soft. From 0.6, the butterfly clip and white lashes lock in. At 0.8, E5 / E7 keep the collar trim most cleanly; 1.0 adds little beyond hair bloom.

Full body and outfit control

Black-dress full body	Casual swap attempt

Black-dress full-body shots keep identity from 0.4; skirt striping is clearer at 0.6–0.8. Casual prompts without the outfit token still usually return the black dress — expected for a single-outfit set.

Scene generalization

Identity holds in the night-forest tests. From 0.8, E5 / E7 keep more dress detail without collapsing into a single training still.

Default-outfit leak

With only echidna_rezero and no clothing words, the default high-collar black dress appears. Other clothes must be written in the prompt.

Trigger isolation

With the LoRA loaded but without echidna_rezero, the controls are ordinary dark-haired characters: no white hair, white lashes, or green butterfly clip. The character binding stays on the trigger.

Training

• Data: 25 manually selected Echidna stills with matching .txt captions (5 holdout images reserved locally, not published). Captions start with echidna_rezero, echidna_black_dress, keep_tokens = 2, remaining tags shuffled. Published as abbuibuibui/echidna-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 7 per image, 7 epochs, 1225 optimizer steps (175 per epoch). One weight file saved per epoch.

Related repositories

• Dataset: abbuibuibui/echidna-Dataset — the 25 image/caption pairs used to train this LoRA.

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.

This is a character LoRA, not a general style model. Evaluation used waiIllustriousSDXL v170, Euler a, and 832×1216. Other bases, samplers, prompts, or sizes can change the result. Outfit changes are harder than multi-outfit LoRAs because the training set is a single black dress.

License

• Character and source-work rights remain with © Tappei Nagatsuki / White Fox / 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.