【中文】

触手概念 LoRA，训练于 3,000 张按材质分层抽样的图像，专注「材质 / 颜色 / 交互」三个维度的精确控制。

**训练基础**

- 底模：Anima base v1.0（Cosmos-Predict2-2B + Qwen3-0.6B 文本编码器 + Qwen-Image VAE）

- 训练量：3,000 张 / 1 epoch / 3,000 步 / 89 分钟（RTX 4060 Laptop 8GB）

- network_dim 32 / network_alpha 16 / T-LoRA（min_rank 4）

**实测能力（106 张对比图验证）**

材质 8 种，6/8 完全符合：

光滑橡胶（镜面高光 + 水珠）、哑光天鹅绒、透明玻璃（真的能看到背后人物）、生物荧光（自发光照亮皮肤）、金属分节（镀铬镜面 + 拼缝）、黏液凝胶（滴落黏液丝）、粗糙岩石、肌肉肉质

颜色 9 种，准确率 100%：

深紫、猩红、翠绿、淡青、漆黑、象牙白、橙金、樱粉、灰蓝

交互 8 类，6/8 成功：

抓手腕、缠腰、缠脚踝拖拽、塞入口中、环颈、背后抱住、抬起下巴、全身包裹

**使用要点**

提示词里的触手必须有「接触动作」，写成悬垂背景会画不出人物：

✅ tentacles coiling around her waist

❌ tentacles hang from the canopy

推荐前缀：masterpiece, best quality, full body shot of an anime girl...

**已知短板**

- full body shot 不是 100% 可靠，偶尔裁掉头部

- 红/绿/橙三色容易把触手形态带塌成肉块或裙摆，建议加 with clear round suckers and coiling tips 锁定形态

- 描述 muscle / veins 时容易污染人物皮肤

**推荐参数**

LoRA 权重 1.0 / 采样步数 28 / CFG 4.0 / 分辨率 1024×1024 / euler + simple

---

【English】

Tentacle concept LoRA trained on 3,000 images stratified by material, focused on precise control across three dimensions: material, color, and interaction.

**Training basis**

- Base model: Anima base v1.0 (Cosmos-Predict2-2B + Qwen3-0.6B text encoder + Qwen-Image VAE)

- Scale: 3,000 images / 1 epoch / 3,000 steps / 89 minutes (RTX 4060 Laptop 8GB)

- network_dim 32 / network_alpha 16 / T-LoRA (min_rank 4)

**Verified capabilities (106 comparison images)**

Materials — 8 types, 6/8 fully accurate:

glossy rubber (specular highlights + water droplets), matte velvet, transparent glass (you can genuinely see the body behind it), bioluminescent (self-illuminating, lights up her skin), segmented chrome (mirror finish + panel seams), slime gel (dripping viscous strands), rough rock, fleshy muscular

Colors — 9 types, 100% accurate:

deep purple, crimson, emerald green, pale cyan, pitch black, ivory white, golden orange, sakura pink, slate blue-grey

Interactions — 8 types, 6/8 successful:

gripping both wrists, wrapping the waist, coiling ankles and pulling, inserted into mouth, looped around neck, embracing from behind, lifting the chin, full-body cocooning

**Usage notes**

Tentacles in your prompt must have a contact action. Writing them as hanging background makes the model drop the character entirely:

✅ tentacles coiling around her waist

❌ tentacles hang from the canopy

Recommended prefix: masterpiece, best quality, full body shot of an anime girl...

**Known limitations**

- "full body shot" is not 100% reliable — occasionally crops the head

- Red/green/orange tend to collapse the tentacle form into slabs or skirt shapes; add "with clear round suckers and coiling tips" to lock the shape

- Describing "muscle"/"veins" can bleed onto the character's own skin

**Recommended settings**

LoRA weight 1.0 / 28 steps / CFG 4.0 / 1024×1024 / euler + simple