wesjos/Z-image-turbo-silverwolf-Lora

Silver Wolf v5 (ai-toolkit trained by my Ai agent fairy) - Honkai: Star Rail

Character LoRA for Z-Image-Turbo (8-step distilled diffusion model by Tongyi MAI / Comfy-Org). Trained on the official Silver Wolf (银狼) from Honkai: Star Rail (崩坏:星穹铁道).

This is a character-identity LoRA for the cyberpunk hacker girl Silver Wolf. Trained with ostris/ai-toolkit + the official zimage_turbo_training_adapter_v2.safetensors which preserves the 8-step distilled trajectory of the base model.

Trigger Word

silver wolf from honkai star rail, 1girl

Prompt sampple

cute anime girl, silver hair, wearing purple t-shit and white short, standing waving hello，white background

Important: Do NOT use bare silver wolf — base Z-Image will interpret it as a literal wolf. Always include the full from honkai star rail anchor.

Recommended Inference Settings

Parameter	Value	Notes
LoRA strength	0.85	SSOT golden value; sweet spot (base 4.0 → LoRA 7.5)
Steps	12	Z-Image-Turbo 8-step distilled + 4 extra for quality
CFG	1.0	Recommended for distilled Turbo models
Sampler	flowmatch	Native to Z-Image-Turbo
Scheduler	simple	Default
Shift	3.0	AuraFlow shift (do not change)
Resolution	1024×1024 or 1920×1080	Z-Image-Turbo native; 16:9 for wallpapers

Training Details

Item	Value
Base model	Comfy-Org/z_image_turbo (8-step distilled, INT8 ConvRot)
Training framework	ostris/ai-toolkit (official Z-Image-Turbo trainer)
Training adapter	ostris/zimage_turbo_training_adapter_v2.safetensors (preserves 8-step distillation)
Architecture	arch: zimage (NOT z_image)
Quantization	quantize: true + quantize_te: true (dual float8)
Dataset	500 images, Qwen-VL B 档 relabeled to Danbooru-style tags
Trigger word	silver wolf from honkai star rail, 1girl
Network	rank=32, alpha=32, linear LoRA
Optimizer	adamw, lr 1e-4
EMA	use_ema: true, ema_decay: 0.99
Batch size	1, gradient_accumulation 1
Steps	3000 (≈ 6 epochs on 500 imgs)
Resolution	1024×1024
Precision	bf16
Training time	~2.1 hours on RTX 4090 24GB
VRAM	~16.6 GB / 24 GB

Validation Results

4-tier verification (主人 9-3 钦定常驻规则):

1. md5: all pairwise different → LoRA actually applied
2. Pixel diff vs base: 0.85 gives 33% / 98% vs base → LoRA pushes real weight
3. vision_analyze 9-02 钦定 zero-tolerance check: 2 legs / 2 feet / 2 arms / 5+5 toes ✓
4. Mean value monotonicity with scale: 202→207→213→230→237

Subjective scoring (vision_analyze 1-10):

• Base: 4.0/10
• LoRA @ 0.85: 7.5/10 (+3.5 lift)
• LoRA @ 1.0: 7.0/10 (push heavy → local details crash)

Training Samples

Sample generated during training (every 500 steps), scale=0.85, 9 steps, prompt: silver wolf from honkai star rail, 1girl, standing, looking back over shoulder, barefoot, soles, toes, white background

Step	Sample
0	noise initialization
500	early learning, basic composition
1000	character features emerging
1500	outfit + hair details solidifying
2000	fine details converging
2500	near-final quality
3000	final, deployed version

Files

• silverwolf_v5_aitk.safetensors (170 MB) — the LoRA weights (480 keys, diffusers format)
• config.yaml — training configuration
• optimizer.pt — optimizer state (for resume)
• samples/ — training-progress samples

ComfyUI Usage

# Place at:
/mnt/sdb/ubuntu/ComfyUI-master/models/loras/silverwolf_v5_aitk.safetensors

# Use via LoraLoaderModelOnly node:
#   lora_name: silverwolf_v5_aitk
#   strength_model: 0.85

Or via API:

python /tmp/comfyui_t2i_lora_api.py "silver wolf from honkai star rail, 1girl, \
  standing, looking back over shoulder, barefoot, soles, toes, white background" \
  --lora silverwolf_v5_aitk --lora_strength 0.85 --steps 12 --cfg 1.0

Known Limitations

• NSFW / foot fetish content: hand-check by the user, do NOT use vision_analyze (platform blocks)
• Risk postures (prone/suspended/reclining + LoRA scale ≥ 0.5) can produce wrong toe count — recommend thigh high boots to avoid
• Trigger ambiguity: bare silver wolf triggers base model's animal distribution — always include from honkai star rail
• Base dependency: requires Z-Image-Turbo with INT8 ConvRot weights, won't work on raw bf16 base
• 6 epochs may be tight for some aesthetics — if character drifts, consider retraining with rank 64

Companion LoRA: Hoyoverse (Style Filter)

For a cyberpunk purple-black style filter that works on any ACG character (incl. crossover), see the companion Hoyoverse LoRA — different purpose (style filter, not character).

License

This LoRA is released under fair-use-academic-noncommercial:

• ✅ Personal use, research, art creation
• ✅ Open-source / non-commercial derivative work (with attribution)
• ❌ Commercial use without permission from the rights holders of Honkai: Star Rail
• ❌ Misrepresentation of the trained character as your own original creation

Character Silver Wolf (银狼) is © COGNOSPHERE / HOYO-MIX / miHoYo. This LoRA is a fan-made derivative for research and personal artistic expression.

Citation

@misc{silverwolf-v5-aitk,
  author = {wesjos},
  title = {Silver Wolf v5 Z-Image-Turbo LoRA (Honkai: Star Rail)},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/wesjos/silverwolf-v5-aitk}}
}

Acknowledgements

• Base model: Comfy-Org/z_image_turbo by Tongyi MAI
• Training framework: ostris/ai-toolkit by ostris
• Training adapter: ostris/zimage_turbo_training_adapter (preserves 8-step distillation)
• Vision language model: Qwen3.8-27B by Alibaba (used for caption relabeling)
• Character: Silver Wolf (银狼) from Honkai: Star Rail by COGNOSPHERE / miHoYo