极简 Q 版头像风格 LoRA。大圆头被画框裁切、两只近黑竖胶囊眼、无嘴无鼻、
扁平色块、纯黑背景。适合头像、表情包、图标。
触发词：chibi_head
四个角度标签：peek from bottom left / peek from bottom right / front view /
strong head tilt
参数：MOSS-Image Base，CFG 3，Shift 5，1024×1024，20 到 50 步，euler + simple。
只画头，不画身体和场景。耳饰、蕾丝这类小细节画不出来。
moss底模原生也是可以生成出来这个Q版头像，但是需要很一大段的提示词。而且不一定满意。所以我训练了这个lora加强了一下Q版头像的细节。
----
## What this is
A style LoRA for minimal 2D chibi head icons. Large round head cropped by the
frame, two near-black vertical capsule eyes, no mouth, no nose, soft oval blush,
flat color blocks, pure black background. Built for avatars, stickers and
app icons.
## Trigger word
chibi_head
Always include it. Everything else is optional.
## Angle tags
The dataset was built with four compositions, each with its own tag:
| Tag | Composition | Share |
|---|---|---|
| peek from bottom left | head enters from lower left, tilted clockwise | 22 |
| peek from bottom right | head enters from lower right, tilted counter-clockwise | 12 |
| front view | head low and centered, nearly upright | 12 |
| strong head tilt | 28 degree tilt, clearly slanted | 5 |
Leave the tag out and the model picks a composition on its own. Keep in mind
strong head tilt had the smallest share in training, so it is the least
reliable of the four.
## Recommended settings
| Setting | Value |
|---|---|
| Base model | MOSS-Image Base (also works on SFT) |
| Steps | 20 – 50 |
| CFG | 3.0 |
| Resolution | 1024 × 1024 |
| Sampler | euler |
| Scheduler | simple |
Shift matters. This is a flow matching model; running it without shift 5 will
produce noise.
## Prompting
Keep it short and put the features first:
chibi_head, blonde hair, red wolf ear hood, peek from bottom left
Anything you want visible has to be named — hair color, ears, horns, hats,
headpieces. The style does not invent accessories.
## What it does well
- Character heads with strong silhouettes: animal ears, horns, hoods, hats,
halos, large headpieces
- Dark or near-black backgrounds, flat vector look, sticker and icon output
- Consistent framing across a set of characters
## What it does not do
- Full body, hands, or any scene beyond the head
- Realistic rendering, 3D, painterly texture
- Fine detail: jewelry, lace, individual hair strands, small ornaments
- Very dark hair on the pure black background — the silhouette can merge into
the backdrop. Use a hair tone one step lighter than black if this happens.
## Notes
If your output looks like colored static, check shift and CFG first. Those two
account for almost every failure case reported so far.