Krea2 Pussy LoRA. Trained on the same dataset as: Z-Image - Better Pussy & Poses
Workflow (krea2_simple_worflow.json) attached to the optional files.
No need to use conditioning node or safety filter bypass lora！
I feel like Krea2 LoRA training is much easier than Z-Image. I got good pussy quality with just 2,000 steps.
Base model (Krea2 Turbo): https://huggingface.co/Comfy-Org/Krea-2/blob/main/diffusion_models/krea2_turbo_fp8_scaled.safetensors
Text encoder: https://huggingface.co/Comfy-Org/Krea-2/blob/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors
Vae: https://huggingface.co/Comfy-Org/Krea-2/blob/main/vae/qwen_image_vae.safetensors
Dataset: 311 unique pussy images (163 close-up; 20 median close-up; 127 full-body)
Steps: 2000
Pussy lora strength: 1.0 - 1.5
CFG: 1.0
Steps: 8
Sampler: euler
Scheduler: simple
Resolutions: 1024/1536/2048
⚠️ This LoRA was trained on Chinese captions. English might work since Qwen3VL is bilingual
Required trigger word: anatomy_detail
Subject: 东亚年轻女性 (young East Asian woman)
Control Tags:
Framing: close-up / medium / full body
Labia: 阴唇闭合 / 阴唇被掰开内部可见 (labia closed / labia spread open, inside visible)
Pubic hair: 阴毛可见 / 没有阴毛 (pubic hair visible / hairless_pussy)
Legs & pose: 双腿打开 / M字腿膝盖弯曲 / V字腿膝盖伸直 / 站立一字马 / 横向一字马 (legs spread / M-legs knees bent / V-legs knees straight / standing split / horizontal split )
Hands: 双手绕过屁股外侧掰开阴部 / 双手放置于下腹部掰开阴部 (hands around butt spreading genitals / hands on lower abdomen spreading genitals)
Views: front / side / back view
prompt = “anatomy_detail” + framing + subject + labia state + pubic hair state + legs + hands + view + other details
Sample prompt：
anatomy detail, full body, East Asian young woman, closed labia, visible pubic hair, legs spread apart, M-legs knees straight, hands around butt spreading genitals, front view, lying on a red plastic track, eyes gazing directly at the camera, sneakers, sportswear, short sports skirt, ponytailDescription
Initial version trained on Krea2 Raw
Trained steps: 2000
Rank: 32
Optimizer: AdamW
Initial learning rate: 0.0001
lr_scheduler: constant
Trained resolutions: 1024/1536/2048 (Use these for generation too)
FAQ
Comments (2)
Hi my friend, this works very cleanly with my setup. I was surprised at its ability to maintain my character lora likeness. Just a question out of curiosity, did you include faces in dataset?
Dataset includes 163 close-ups, 20 medium close-ups, and 127 full-body shots. Only the full-body images have faces, and all of them are swapped to a face I like using the Face Detailer node. However, it seems face converge slower than pussy for Krea2. I feel, for 2000 steps, this LoRA won't affect the character's face too much.