wesjos/Qwen-Image2.1-ZenlessZoneZero-Lora

ZZZ Style LoRA for Qwen-Image-2.1

A Zenless Zone Zero-inspired style LoRA trained on top of Qwen-Image-2.1. The trigger word is zenlesszonezero (optional — the style also transfers without it).

---

🎨 Gallery

_{All six images are 1024×1024 and were generated with this LoRA. Generated content is not necessarily typical of the source material in every respect — see [Limitations](#-limitations).}

---

📋 Model Details

Property	Value
Architecture	Qwen-Image-2.1 (qwen_image_2)
Type	Style LoRA (character-expressive / hand-feel optimisation)
Trigger word	zenlesszonezero(1.6)/zzz_style(1.4) — optional
Rank / Alpha	32 / 64
Base model quantisation	Comfy-Org int8 convrot + convrot8 training
Training framework	ai-toolkit (flowmatch / shift)
Default LoRA file	zzz_v1.4_step3000.safetensors

---

📦 Checkpoint Guide

File	Recommendation
zzz_v1.4_step1000.safetensors	Under-trained — the style is not yet fully formed
zzz_v1.4_step2000.safetensors	✅ Recommended — style is complete without over-fitting
zzz_v1.4_step3000.safetensors	✅ Recommended — the default; style is complete without over-fitting
zzz_v1.4_step4000.safetensors	Slightly over-fitted

---

🚀 Usage

Prompt template

zzz_style, furina from genshin impact, waving hands

Recommended LoRA scale

Scale	Effect
0.6 – 0.8	Light style, more freedom for the content
0.8	Sweet spot — full style without crushing the composition
1.0 – 1.1	Strongest style, may start to override composition / pose

---

ComfyUI

Base model files (diffusion_models + text_encoder + vae):

• qwen_image_2.1_int8_convrot.safetensors — diffusion model
• qwen3vl_8b_int8_convrot.safetensors — text encoder
• qwen_image_2.1_vae_bf16.safetensors — VAE

Sampler settings (KSampler):

Parameter	Value
Steps	25
CFG	4.3
Scheduler	euler + simple
Shift	3.1

---

Negative prompt

None required — leave it empty. If quality wobbles, try blurry, low quality, jpeg artifacts.

---

⚠️ Limitations

1. The training set is dominated by single-girl portraits, so multi-character scenes and male characters generalise poorly.
2. The base model used for training was int8 quantised (convrot8).
3. Generated content may not be typical of the source material in every respect.

---

📄 License & Attribution

Released under Apache-2.0.

• Base model — Qwen-Image-2.1 (Alibaba Qwen team)
• Training framework — ai-toolkit (ostris)
• Art style reference — the art team of Zenless Zone Zero by miHoYo

Zenless Zone Zero is a trademark of miHoYo / HoYoverse. This LoRA is a third-party fan work, not affiliated with or endorsed by miHoYo.