This is part 2 of a previous article related to Qwen-Image NSFW and LoRA characters:
https://civitai.com/articles/18998/basic-guide-to-qwen-image-lora-training
https://civitai.com/articles/18798/qwen-image-nsfw-lora-notes
Please keep in mind that these are my personal experiences, primarily shaped by trial and error. The purpose is to give you a head start, helping you avoid my mistakes or build upon them.
Demo images related to the training settings are in a ZIP file (NSFW).
Images for the character are sourced from: https://civitai.com/user/squishypotato/images
Step 1: Build Your Dataset
The dataset is the heart of your LoRA! it’s what teaches the model to generate hot, accurate poses like missionary or full nelson. We need a tight, high-quality set of images with clear captions so the model learns exactly what you want. Here’s how to do it, no fluff.
How Many Images:
Start with 30 - 40 NSFW images. (NSFW + 10 to 20 neutral poses, refer to first part.)This is enough to train a focused LoRA without wasting time. You can add more later if needed, but don’t go overboard (in case of character, over 100 risks the model getting stuck on repeats).
Image Specs:
Size: At least 1024x1024. Square or close to it works best.
Format: PNG or JPG, high quality. No blurry, pixelated crap.
Style: Mostly photorealistic (like real couple photos, 70%) with a few semi-stylized (like sexy digital art, 30%) if you want variety. Avoid cartoonish stuff unless that’s your goal.
No Distortions: Pick images with clean anatomy. No extra or missing fingers non-visible arms, weird legs, fingers, or funky faces. The model learns from what you give it, so garbage in = garbage out.
Captions: Every image needs a short, clear caption (1 sentence) describing the pose, who’s involved, and the vibe. Save captions as text files (e.g.,
image001.txtnext toimage001.png) or in a singledataset.csvdepending of your training method, software ect.:Example: “Man and woman in missionary, her legs wrapped around him, dim bedroom light, super realistic.”
Example: “Couple in doggy style, her hands gripping sheets, him behind, moody red lighting, photorealistic.”
Don’t use vague stuff like “sexy pose” or “hot couple.” Be specific: name the pose (missionary, doggy, etc.), mention body positions, and add a vibe (romantic, intense, etc.).
You need to specifically describe what is visible. If you don't mention that the rectum is visible, there is a 90% chance it won't show up. I learned this through at least five training sessions. I did everything by the book, yet it didn't show up. [Personal experience]
Sometimes, using slang terms for sexual acts in Chinese can play a role. (I asked AI for the top slang names for vagina, penis, etc, and tried them in my prompt. It made a difference.again, personal experience.)Read the prompt (https://civitai.com/images/99098288) now ,the workflow is there. try to achieve it in English.
- Where to Get Images:
- Find Them: Use CivitAI (Pay close attention to fingers, scale, etc. great majority of the materials are useless for high quality LoRA. Remember, your eye edits the photo for you. Zoom in on fingers, eyes, etc. AI can see those small deformed fingernails, and you might think your model is underfitting or overfitting ), stock photo site like Unsplash or Pexels with searches like “romantic couple photorealistic” or “intimate couple photography.” You might not find exact poses like full nelson (stock sites are tame), so grb 10 to15 close matches (like couples hugging or lying close) or Barve Browser -> Private Window with Tor -> Duck.com or Yandex image, and search.
Generate Them: Use SDXL eco-system to create the spicier poses. Example prompt: “1girl, 1boy, missionary, legs_up, on_back, male_on_top, bedroom, soft_lighting, photorealistic, highres, detailed_skin, realistic_anatomy,” Generate 5 to10 per pose, then pick the best 4 to 6.
Mix It Up: Aim for 50% couple poses (missionary, doggy, etc.) and 50% solo poses (example: someone lying back or leaning forward) to teach the model body mechanics first.
What Poses to Include
Here’s a clear list of poses and variations to include in your 30 - 40 images. I’ve split it into categories (missionary, doggy style, full nelson) plus some extras to make the model flexible. Aim for 4 to 6 images per category to hit your total. These are described in a way that’s natural but keeps it artistic for training purposes.
1. Missionary Poses (6 - 8 images):
Classic missionary: Woman lying on her back, legs up or wrapped around man’s waist, him above.
Caption: “Man and woman in missionary, her legs around him, intimate bedroom, realistic. [mention what's visible, breasts, testicles, etc]”
Variations: Legs bent at knees (2 images), legs spread wider (2 images), one leg over his shoulder (2 images).
Angles: Eye-level (50%), overhead (30%), side view (20%). Add soft lighting or candles for mood.
2. Doggy Style Poses (6 - 8 images):
Standard doggy: Woman on all fours, man behind, hands on her hips or sheets.
Caption: “Couple in doggy style, her hands gripping bed, him behind, dim room [mention what is visible, dick, pussy etc,].”
Variations: Woman’s upper body lower (2 images), arched back (2 images), one hand reaching back (2 images).
Angles: Side view (50%), back view (30%), slightly angled (20%). Try moody or warm lighting.
3. Full Nelson Poses (4–6 images):
Full nelson: Man standing or kneeling, woman’s legs up and arms pinned back, him supporting her.
Caption: “Couple in full nelson, her legs lifted, him behind, intense vibe, realistic.”
Variations: Standing (2 images), sitting on edge of bed (2 images), slight tilt in her body (2 images).
Angles: Side view (60%), three-quarter view (40%). Use dramatic lighting (like low light with shadows).
4. Other Intimate Poses (6 - 8 images):
Spooning: Both lying on their side, him behind, her curled into him.
Caption: “Couple spooning, his arm around her waist, cozy bedroom, photorealistic.”
Cowgirl: Woman on top, sitting or leaning forward, man lying back.
Caption: “Woman on top in cowgirl, hands on his chest, soft light, realistic.” (don't forget about reverse cowgirl)
Standing close: Face-to-face, bodies pressed together, hands on waists or faces.
Caption: “Couple standing close, her hands on his neck, his on her hips, romantic vibe.”
Variations: Mix angles (front, side, back) and settings (bedroom, living room, or outdoor like a beach).
5. Solo Poses for Anatomy (4 - 6 images):
Woman lying back: Legs bent or spread, arms relaxed, prepping for missionary-like poses.
Caption: “Woman lying on bed, legs slightly up, soft lighting, realistic.”
Man leaning forward: Hands positioned as if gripping hips, for doggy or full nelson prep.
Caption: “Man leaning forward, hands out, intense vibe, photorealistic.”
Variations: Different body angles (lying, kneeling, standing) to teach limb placement.
Total Breakdown for 35 Images:
| Category | Images | Focus |
Missionary | 8 | Legs, closeness, upper body.
Doggy Style | 8 | Back arch, hand placement .
Full Nelson | 6 | Lifted legs, arm pinning.
Other Intimate | 8 | Spooning, cowgirl, standing close.
Solo Poses | 5 | Anatomy basics for each pose.
Total | 35 | Angles: 40% side, 30% front, 20% back, 10% overhead.
Diversity Tips:
People: Mix body types and skin tones (e.g., 20% variation in ethnicity/age 18+). If it’s couples, keep it 50/50 male/female or adjust to your preference.
Lighting: Use soft bedroom lights (50%), warm candles (30%), or dramatic low light (20%) to add mood.
Backgrounds: Bedrooms (50%), living rooms (30%), outdoor like parks/beaches (20%) to make the model adaptable.
Vantage Height:
Knee Level.
Eye-Level.
High Angle.
Low Angle.
Bird's Eye View.
Worm's Eye View.
Shoulder Level.
I ran a few test sessions:
LR 0.00005, 80 repeats per image => underfitting (although the safest LR in my opinion, but it takes quite a long time).
LR 0.0004, 50 repeats per image => overfitting.
LR 0.0003, 88 repeats per image => about 10% underfitting. In some poses, genitalia for both male and female were greatly underfit. My dataset might have contributed to this (post training inspection of dataset).
Demo (LR 0.0003):
I haven't yet tried LR 0.0003 with 100 repeats per image. If I do, I'll update the article.
There you have it.
Good luck!