Zenless Zone Zero Chibi LoRA — Qwen-Image-2.1
note: base model is actually from ComfyUI-org int8 version. trigger
zenlesszonezero_chibi.
🎨 Gallery
st3000 · s0.70 |
st2500 · s0.65 |
st3000 · s0.60 |
st2500 · s0.70 |
st3000 · s0.90 |
st2500 · s0.70 |
st2500 · s0.70 |
st2500 · s0.70 |
st2500 · s0.70 |
📋 Model Details
| Property | Value |
|---|---|
| Base model | Qwen-Image-2.1 (qwen_image_2) |
| Type | Style LoRA — chibi / Q-version proportion compression |
| Trigger word | zenlesszonezero_chibi — optional; the style can also transfer naturally |
| Rank | 16 |
| Tensors | 384 (192 pairs of lora_A/lora_B) |
| Target modules | Per layer: attn.to_q · to_k · to_v · to_out.0 · img_mlp.gate_up · img_mlp.out |
| Blocks covered | All 32 layers transformer_blocks.0–31 (no missing, no extra layers) |
| Dtype | F16 |
| Alpha | Not written → ComfyUI / diffusers treat as alpha = rank = 16 (i.e., scale 1.0) |
| Training framework | ai-toolkit v0.13.21 · flowmatch |
| Trained on | int8 convrot quantized base model (Comfy-Org) |
| Size per file | 76 MiB |
📦 Checkpoint Guide
| File | Step / Epoch | Recommendation |
|---|---|---|
zzz_chibi_v2_st2000.safetensors |
2000 / 8 | ✅ Safe bet; style is formed, least underfitting |
zzz_chibi_v2_st2500.safetensors |
2500 / 11 | ⭐ Recommended default — most example images come from this checkpoint |
zzz_chibi_v2_st3000.safetensors |
3000 / 13 | ✅ Strongest style; chibi proportions are flatter and more "sticker-like" |
🚀 Usage
Prompt
zenlesszonezero_chibi, nahida, shy
zenlesszonezero_chibi, a girl with fox tail, shy
Both English and Chinese prompts work; character names can be written directly (nahida / furina / 不知火舞 have all been tested).
LoRA strength
| Scale | Effect |
|---|---|
| 0.55 – 0.65 | Mild chibi; retains more original detail and compositional freedom |
| 0.70 | Empirical sweet spot — chibi features are clear, image is not overly compressed |
| 0.80 – 0.90 | Stronger chibi compression; head-to-body ratio becomes more extreme; above 1.0 composition starts to suffer |
07/08are strictly controlled strength comparison references: same prompt, same checkpoint (st2500), same strength 0.70, seeds differ by only 1, allowing visual comparison of randomness.04/05share a prompt but differ in seed, checkpoint, and strength — not a strict strength comparison experiment.
ComfyUI
Base model files (diffusion_models + text_encoder + vae):
| Slot | File |
|---|---|
| diffusion model | qwen_image_2.1_int8_convrot.safetensors |
| text encoder | qwen3vl_8b_int8_convrot.safetensors |
| VAE | qwen_image_2.1_vae_bf16.safetensors |
KSampler parameters (measured values from this repo's example images):
| Parameter | Value |
|---|---|
| Steps | 25 |
| CFG | 5.0 |
| Sampler | euler |
| Scheduler | simple |
| Shift (ModelSamplingAuraFlow) | 3.0 |
| Size | 1024 × 1024 |
Negative prompt
deformed, wrong hands, mutated, fused fingers, extra hand, extra feet, ugly underwear
⚠️ Limitations
- Training set is mainly single-character half-body / portrait shots — multi-character scenes, male characters, and full-body compositions generalize poorly.
- Trained on an int8 quantized base model — inherits some of convrot8's texture loss; detail sharpness is slightly lower than fp16 training.
- Chibi style actively simplifies — complex accessories and multi-layered clothing patterns are easily flattened; lower strength below 0.6 when detail matters.
- Generated content may not match official lore — character details and clothing patterns should defer to official source material.