lantianws666/moss-chenbin-style

Chenbin Style | 鬼针草画风

Chenbin Style (鬼针草画风)

还原画师鬼针草标志性的柔和细线条、低饱和水彩质感、透亮肌肤与淡樱粉腮红。 A style LoRA recreating the soft linework, low-saturation watercolor texture, translucent skin and signature pale blush of artist Chenbin.

Recommended Settings

ParameterValueLoRA Weight0.8 (0.6 – 1.0)Steps50CFG Scale3.0SamplerEulerSchedulerSimpleResolution1024×1024 – 1536×2048Skip Layer Guidancesingle_layers 22, start 0.0, end 1.0

Clip skip is not used. No trigger words required — this is a pure style LoRA.

Training Details

ItemValueBase ModelOpenMOSS-Team / MOSS-Image-BaseNetworkLoRARank / Alpha32 / 32Steps1500Batch Size1Learning Rate1e-4OptimizerAdamW 8bitPrecisionbf16Text Encoder TrainedNoResolution1024×1024Dataset117 imagesuncond_skip_layers[22]Training Time~55 min (H20 96GB)Training Toolai-toolkit

Usage Notes

No trigger words — the style applies automatically

Weight 0.8 is the sweet spot. 1.0 tends to overexpose, below 0.65 the style fades

Keep CFG at 3.0. Going above 5 causes oversaturation

Steps below 30 visibly loses detail in hair strands and lace

SLG must stay on (single_layers 22). Turning it off produces color noise

When stacking with a character LoRA, use 0.5–0.6 for each

Known Limitations

Fine textures — trained at 1024px, so ultra-fine details like stocking mesh cannot be rendered. This matches the artist's loose, impressionistic style.

White-heavy bias — the training set skews toward white clothing, so pure-white scenes may overexpose.

Auto pantyhose — the model tends to add pantyhose automatically. Use negative prompts to remove.

License

Commercial use: No

Merging: Allowed

Selling generated images: Allowed

Credit appreciated but not required

Tags anime, style, illustration, watercolor, soft, pastel, low saturation, fine line, blush, gentle, 鬼针草, chenbin

Files

File	Size	Note
chenbin_style_v2_blush_1500step_rank32.safetensors	141.8 MB	bf16 safetensors, rank 32

Trigger word: 无（纯画风，自动生效）

Recommended LoRA weight: 0.8

Samples

Usage

This is a LoRA for MOSS-Image. To use it on HuggingFace Spaces, pass this repo id to the lora_repo field of mrfakename/moss-image:

lora_repo  = "lantianws666/moss-chenbin-style"
lora_scale = 0.8
model      = "MOSS-Image Base"

Or load it with the native MOSSImagePipeline:

mc = ModelConfig(
    name_or_path="OpenMOSS-Team/MOSS-Image-Base",
    arch="moss_image", dtype="bf16",
    quantize=False, quantize_te=False,
    model_kwargs={"uncond_skip_layers": [22]},
)
lora = InferenceLoRA("chenbin_style_v2_blush_1500step_rank32.safetensors", strength=0.8)

Recommended settings

Item	Value
Base model	MOSS-Image
LoRA weight	0.8
Resolution	1024x1536
Steps	30
CFG	3.0
uncond_skip_layers	[22]

Training

Item	Value
Tool	ai-toolkit
Network	LoRA
Rank / Alpha	32 / 32
Text encoder	frozen (not trained)
Precision	bf16

License

Apache-2.0 (weights). See the original Civitai page for usage notes.