One issue I observed in V1.0 and V2.0 was that the labia appeared excessively dark in generated images, unlike their appearance in the training dataset. This likely resulted from overfitting due to my original training approach: I masked out the genital region in full-body images, assigning a loss weight of 1.0 only to the masked area and 0.0 elsewhere.<h3 id="in-v3.0-i-removed-this-masking-strategy-and-added-additional-sexual-pose-variations-to-the-dataset.-hope-you-enjoy-the-improvements!">In V3.0, I removed this masking strategy and added additional sexual pose variations to the dataset. Hope you enjoy the improvements!</h3><h3 id="v1.0-dataset-contained-only-solo-female-images-causing-conflicts-with-other-hardcore-loras-and-checkpoints.">V1.0 dataset contained only solo female images, causing conflicts with other hardcore LoRAs and checkpoints.</h3><h3 id="v2.0-includes-200-images-of-sex-scenes-to-the-dataset-to-improve-compatibility.">V2.0 includes 200 images of sex scenes to the dataset to improve compatibility.</h3><h3 id="lora-strength:-1.0-1.5">lora strength: 1.0 - 1.5</h3><h3 id="if-mosaics-appear-increase-the-lora-strength.">If mosaics appear, increase the LoRA strength.</h3>Sex positions included:<ul><li>Reverse Cowgirl</li><li>Missionary</li><li>Doggy Style</li></ul>Krea2 Pussy LoRA. Trained on the same dataset as: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2723055/z-image-better-pussy-and-poses">Z-Image - Better Pussy &amp; Poses</a> Workflow (krea2_simple_worflow.json) attached to the optional files.<h3 id="no-need-to-use-conditioning-node-or-safety-filter-bypass-lora">No need to use conditioning node or safety filter bypass lora！</h3>I feel like Krea2 LoRA training is much easier than Z-Image. I got good pussy quality with just 2,000 steps.<ul><li>Base model (Krea2 Turbo): <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Krea-2/blob/main/diffusion_models/krea2_turbo_fp8_scaled.safetensors">https://huggingface.co/Comfy-Org/Krea-2/blob/main/diffusion_models/krea2_turbo_fp8_scaled.safetensors</a></li><li>Text encoder: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Krea-2/blob/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors">https://huggingface.co/Comfy-Org/Krea-2/blob/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors</a></li><li>Vae: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Krea-2/blob/main/vae/qwen_image_vae.safetensors">https://huggingface.co/Comfy-Org/Krea-2/blob/main/vae/qwen_image_vae.safetensors</a> </li></ul>Dataset: 311 unique pussy images (163 close-up; 20 median close-up; 127 full-body)Steps: 2000<pre><code>Pussy lora strength: 1.0 - 1.5
CFG: 1.0
Steps: 8
Sampler: euler
Scheduler: simple
Resolutions: 1024/1536/2048</code></pre> ⚠️ This LoRA was trained on Chinese captions. English might work since Qwen3VL is bilingualRequired trigger word: anatomy_detailSubject: 东亚年轻女性 (young East Asian woman)Control Tags:<ul><li>Framing: close-up / medium / full body</li><li>Labia: 阴唇闭合 / 阴唇被掰开内部可见 (labia closed / labia spread open, inside visible)</li><li>Pubic hair: 阴毛可见 / 没有阴毛 (pubic hair visible / hairless_pussy)</li><li>Legs &amp; 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 )</li><li>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)</li><li>Views: front / side / back view</li></ul>prompt = “anatomy_detail” + framing + subject + labia state + pubic hair state + legs + hands + view + other detailsSample prompt：<pre><code>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, ponytail</code></pre>

Krea2 - Better Pussy & Poses

Initial version trained on Krea2 Raw<ul><li>Trained steps: 2000</li><li>Rank: 32</li><li>Optimizer: AdamW</li><li>Initial learning rate: 0.0001</li><li>lr_scheduler: constant</li><li>Trained resolutions: 1024/1536/2048 (Use these for generation too)</li></ul>

Files with SHA256: 2e592e31a9825433076e592b27d509f20346ea6d6328d7498fd76a1e901a843d