Arekejoker/qwen21-nsfw-pose-skill

What this repo does

qwen21‑nsfw‑pose is a ready‑to‑run “recipe” for generating adult‑oriented images with the Qwen‑Image 2.1 text‑to‑image model. It bundles:

• A ComfyUI workflow that runs the model locally (the qwen_image_2.1_int8_convrot checkpoint).
• An NSFW LoRA (low‑rank adaptation) that adds the desired sexual characteristics.
• A library of 136 3‑D poses (derived from public MakeHuman/three‑js data) so the same character can be re‑posed while keeping the face and clothing.
• A set of prompt templates (the “author‑level prompt structure”) that include the required seven elements and LoRA trigger words.
• Scripts for three use‑cases:
  1. T2I – generate a new NSFW image from text.
  2. Edit – keep an existing character and change its pose.
  3. VNCCS PoseStudio – pick an exact pose from the 136‑pose library.

The repository also provides a step‑by‑step “quick‑start” command that reproduces a reference image from a Civitai showcase, plus all the metadata (prompt, seed, sampler settings, etc.) needed to verify the output.

---

Who might use it

User type	Why it’s useful
AI agents / automation tools (Claude Code, Cursor, OpenClaw, etc.)	The repo ships a SKILL.md that describes how an agent can invoke the scripts, enforce the “boldness‑level” ladder (L0‑L3), and warn users when a weaker configuration is used.
Artists / hobbyists who want a local, reproducible pipeline for adult‑content generation without relying on cloud services.
Researchers exploring pose‑controlled generation or LoRA‑based style transfer in a controlled, open‑source environment.

---

How to get it running (high‑level steps)

1. Install dependencies – follow install/setup.md. You’ll need:
   • Python 3.10+ and the requirements.txt from the repo.
   • ComfyUI (the UI framework for diffusion workflows).
   • The base model checkpoint qwen_image_2.1_int8_convrot (downloaded from Hugging Face Comfy-Org/Qwen-Image-2.1).
   • The NSFW LoRA file qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors (Civitai).
   • The 136‑pose library (automatically copied into ComfyUI by the install script).
2. Run the quick‑start command (example from the README):

   python3 scripts/qwen_pose_test.py \
     --mode t2i \
     --name author_case_shower \
     --prompt "$(cat prompts/taws_v13_shower.txt)" \
     --seed 80964 \
     --steps 50 --cfg 2.5 --sampler er_sde --scheduler beta \
     --width 1152 --height 1728 \
     --lora "qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors" \
     --lora-strength 1.0 \
     --out outputs/author_case/shower.png
   

   This reproduces the reference shower image in about five minutes on a modern GPU.
3. Pick a different workflow:
   • --mode edit to keep an existing character and change its pose.
   • scripts/run_vnccs_ui.py --pose <pose_name> to select any of the 136 predefined poses (list them with --list-poses).
4. Adjust the “boldness level” (L0‑L3) by adding LoRA, changing camera angle, or selecting more provocative poses as described in SKILL.md.

---

What’s inside (folder overview)

Folder / file	Purpose
SKILL.md	Formal description of the agent skill – rules, prompt structure, boldness ladder, and required warnings.
install/setup.md	Step‑by‑step installation of the model, LoRA, ComfyUI nodes, and pose library.
scripts/	Python helpers: qwen_pose_test.py (main generation script), run_vnccs_ui.py (pose selector UI), media_lock.py, validate_package.py.
workflows/	JSON workflow definitions for ComfyUI (T2I, edit, VNCCS PoseStudio).
prompts/	The three author‑provided prompt files used in the reference reproductions.
references/	Sample output images that illustrate the L0‑L3 progression.
downloads/samples/	Original Civitai showcase images, plus the JSON metadata (phase5_meta.json) that records every generation parameter.
pose_library/	The 55 hand‑picked pose files (JSON bone rotations) that are merged with the 81 poses from public HuggingFace repos to make the total of 136.

---

Limitations & cautions

• NSFW content only – the LoRA is deliberately tuned for adult imagery; using it for non‑NSFW subjects will produce poor results.
• Legal/ethical – the base model and LoRA are copyrighted by their original creators (HF Comfy-Org/Qwen-Image-2.1 and Civitai authors). The repository does not redistribute them; you must download them yourself.
• Hardware – the recommended settings (1152×1728, 50 steps, er_sde sampler) require a GPU with at least 12 GB VRAM for reasonable speed.
• Prompt discipline – the workflow expects prompts that follow the author’s 7‑element structure (≈300‑1500 characters). Shorter prompts will be rejected or produce low‑quality images.
• Safety – the README warns that the repository is public; viewers who are uncomfortable with explicit images should avoid the references/ and downloads/samples/ folders.

---

TL;DR

qwen21‑nsfw‑pose = a self‑contained, ComfyUI‑based pipeline for generating high‑resolution adult images with the Qwen‑Image 2.1 model, an NSFW LoRA, and a curated 136‑pose library. It offers three modes (text‑to‑image, edit, exact‑pose) and includes a full‑step quick‑start that reproduces a reference Civitai showcase image. Ideal for developers building AI agents that need to invoke a controlled NSFW image‑generation skill, or for hobbyists who want a reproducible local workflow.