训练参数 / Training Parameters
基模型 / Base Model:ideogram-4 (9.3B DiT, FP8)
LuRA 秩 / Rank:64,Alpha:64
数据集 / Dataset:749 张 NSFW 动漫插图 / 749 NSFW anime illustrations
打标格式 / Caption Format:结构化 JSON(high_level_description + style_description + compositional_deconstruction + bbox + color_palette)
分辨率 / Resolutions:512 / 768 / 1024
批次大小 / Batch Size:1
梯度累积 / Gradient Accumulation:2
总步数 / Total Steps:~6000
优化器 / Optimizer:singularity
学习率 / Learning Rate:1e-4
权重衰减 / Weight Decay:1e-4
噪声调度 / Noise Scheduler:Flow Matching
Timestep 类型 / Timestep Type:Linear
Content or Style:Balanced
EMA:开启 / Enabled,衰减 / Decay 0.99
损失函数 / Loss Type:MSE
文本编码器 / Text Encoder:冻结不训练 / Frozen
Caption Dropout:0.05
无翻转增强 / No Flip Augmentation
保存格式 / Save Format:Diffusers,BF16
---
推理建议 / Inference Guide
CFG:3.0–5.0
步数 / Steps:20–30
触发词 / Trigger Word:无 / None
推荐提示词格式 / Recommended Prompt Format:JSON 结构化自然语言描述 / Structured JSON natural language description
风格 / Style: ime NSFWillustration
能力覆盖 / Capabilities:全裸、性行为、体液、多角色交互 / Full nudity, sexual acts, bodily fluids, multi-character
scenes
---
限制 / Limitations
主体为动漫插图风格,写实人像效果不佳 / Anime-illustrationfocused, not sutraits.推荐分辨率 768–1024,过低影响面部细节 / Recommended resolution 768–1024, lower may degrade facial detail.
---
LoRA 权重建议 / LoRA Weight Guide
权重 0.5 / Weight 0.5 — 接近真人风格,皮肤质感偏写实,面部接近照片级 / Approaching photorealistic, skin texture more realistic, facial features near photo-level
权重 0.75 / Weight 0.75 — 真人与动漫混合,过渡自然 / Blended photorealism and anime, smooth transition
权重 1.0 / Weight 1.0 — 纯日系动漫风格,细腻线条,高光皮肤 / Pure anime style, fine linework, glossy skin
权重 >1.0 / Weight >1.0 — 风格过饱和,可补偿基模型的审核抑制,高权重下露骨内容触发率显著提升 / Oversaturated style, can compensate for base model's safety suppression; explicit content triggering rate increases significantly at higher weights
---
声明 / Disclaimer
本 LoRA 仅用于学术研究与艺术创作。请遵守当地法律法规,勿用生成真实人物的
This LoRA is for academic research and artistic creation only. Please couse to generatenon-consensual content of real individuals.





