qwen21-nsfw-pose-skill/README.en.md

qwen21-nsfw-pose

English | 中文

A complete local Qwen-Image 2.1 NSFW / pose pipeline — author-grade prompt structure + NSFW LoRAs + a 136-pose 3D library + precise mannequin pose control with VNCCS.

Built on ComfyUI + the qwen_image_2.1_int8_convrot base kit (the same base model the Civitai authors use). The authors' showcase images were reproduced and verified (2026-10-12).

• For: Claude Code / Codex / Cursor / OpenClaw (agent skill) + manual use
• Includes: runnable author cases, the mandatory recipe, boldness tiers (L0–L3), full dependencies and an install guide

Demo (deform the 3D mannequin → one-click generate → result, Route ④):

---

Showcase

This recipe's output: three tiers of the same seaside scene (L0 → L1 → L2)

L0 · fully clothed, no LoRA (prompt only)	L1 · wet swimwear + LoRA@1.0	L2 · very low angle + wet squat + LoRA@1.0

Three images, same scene, escalating: L0 is prompt-only; L1 adds the LoRA plus explicit skin/water-droplet description; L2 adds the low camera angle and an open pose. The LoRA carries texture, the prompt carries information density, the camera angle carries boldness.

NSFW results (⚠️ the images below contain nudity)

👇 Click the card above to expand the author originals (nudity · 18+, click again to collapse)

Author originals (Civitai showcases — what this recipe targets):

Author · shower (tawsLoRa)	Author · massage (tawsLoRa)	Author · selfie (TheseAlpacas)

The shower image can be reproduced with the command under "Quick start" (the reproduction was verified to match the author original in scene and composition — see downloads/samples/phase5_meta.json and the Quick start section below).

Precise 3D mannequin poses (Route ④, 136-pose library)

Default character image: references/src_standing.png. A pose is a skeleton-rotation JSON (MakeHuman mesh, rendered with three.js), so the pose changes while identity is preserved. Pose library sources: four public HuggingFace repos (Totemistyk/General_Poses, Totemistyk/Ballet_PoseStudio, MIUProject/VNCCS_PoseLibrary_Main, MIUProject/MySuperPoseRepo) plus 55 hand-picked poses — 136 in total.

---

Quick start (follow the author's case)

Install first: see install/setup.md. Once installed, run the self-check (read-only, changes nothing):

python3 scripts/doctor.py

It checks ComfyUI connectivity, base models/LoRAs, VNCCS nodes, the pose library and the browser backend item by item, and prints fix instructions for anything missing. When everything is green, one command reproduces the author's shower image (~5 minutes):

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

Compare against downloads/samples/taws_v13_shower_author.jpg. Full metadata (prompt / parameters / seed / complete workflow): downloads/samples/phase5_meta.json.

Hardware requirements (reference)

The base kit is Qwen-Image 2.1 int8; model files total ~16GB (diffusion 6.8GB / text encoder 8.7GB / VAE 0.6GB), and each LoRA is ≤608MB.

• Minimum: NVIDIA 12GB VRAM + 32GB RAM — it runs, but ComfyUI offloads part of the models to system memory and generation is slower
• Recommended: 24GB VRAM (3090/4090 class) + 32–64GB RAM — models stay resident and it feels smooth
• 8GB VRAM: use a GGUF quantized build (the community reports Q4 runs in 12–16GB VRAM; 8GB with offload is noticeably slower)
• Pure CPU is not recommended (a single image takes far too long)
• Reserve ~40GB of disk; no special CPU requirement (any modern x86_64/ARM will do); Linux is the least fuss; on Windows the base scripts (doctor / --list-poses / media-lock fallback) run natively (community contribution), while the full pipeline is still best on WSL2; macOS (Apple Silicon) has its own section below
• Route ④ additionally needs a desktop browser you can attach to (Playwright/CDP, see install/setup.md)

The showcase images and measurements in this repo come from a unified-memory machine (DGX Spark): 1024×1024, 40 steps, about 32 seconds per image, ~7–10GB peak delta.

macOS (Apple Silicon)

• For M-series chips (M1 and up); Intel Macs are not recommended
• Pick the model by unified memory: ≥32GB, use this repo's original int8 kit directly (ComfyUI + MPS); 16–32GB, prefer a GGUF quantized build (with ComfyUI-GGUF; the community reports Q8 is faster and Q4 saves more memory)
• Speed expectation: minutes per image — community reports: M5 Pro 48GB runs int8 at ~144s per 1024px; M1 Max 64GB at ~7–10 minutes per image (cache-optimizing nodes such as Comfyui-Spectrum-Qwen2.1 can speed this up 2–3×). Compare against this repo's measured 32s per image on DGX Spark, for reference only
• Known issue: edit mode occasionally hits a VAE encoding error on MPS — work around it with ComfyUI's --cpu-vae (ComfyUI issue #16433, a fix is under discussion); when memory is tight, generate one image at a time and close other memory-hungry apps
• Route compatibility: the ①②③ scripts run directly on macOS (pure Python standard library, POSIX-compatible); ④ needs playwright install chromium (the macOS build), and the AppArmor note in install/setup.md does not apply
• ⚠️ The Mac figures above come from community reports. The maintainers have no Mac hardware and have not measured them; if you get it running on a Mac, or hit problems, please open an issue and we will keep this section updated

The four routes

Goal	Route	Command
A new NSFW/sexy image from text	① T2I + NSFW LoRA (mandatory)	scripts/qwen_pose_test.py --mode t2i
Change the pose of an existing character image (keep face/clothes)	② Edit	scripts/qwen_pose_test.py --mode edit
Change pose from a 3D mannequin render (script, no browser)	③ Scripted pose transfer (mannequin reference + VNCCS LoRA)	scripts/run_pose_transfer.py --char <character.png> --pose-ref <mannequin_render.png> --out <result.png>
An exact pose from the 3D library	④ VNCCS PoseStudio	scripts/run_vnccs_ui.py --pose <name> [--prompt "<scene>"]

Route ③'s mannequin reference can be saved automatically by a Route ④ run (the doll_*.png files in the output directory), or any 3D pose render. Do not use a photo or illustration as the pose reference — the model composites the reference person into the frame, or redraws that person entirely; use a textureless 3D mannequin render instead (Routes ③/④).

How to pick a pose (boldness tiers L0–L3, user-request → pose-family mapping): see §4 of SKILL.md, or list all 136 poses with scripts/run_vnccs_ui.py --list-poses.

Recipe highlights (summary)

• er_sde + beta, CFG 2.5, 50 steps, 1152×1728 portrait
• The NSFW LoRA is always on (default femaledVaginus v1.3b @ 1.0)
• Prompts must follow the author's structure: all 7 elements, 300–1500 characters, including the LoRA trigger words
• Weakening the recipe (no LoRA / short prompt / fewer steps) requires telling the user the cost first — never weaken it silently

Full rules and evidence: SKILL.md.

Repository layout

├── SKILL.md            # the agent skill (complete rules)
├── CHANGELOG.md        # release history
├── install/setup.md    # install guide (models / LoRAs / nodes / pose library)
├── scripts/            # doctor.py self-check, generation & custom-pose scripts, browser fallback, media_lock.py, validate_package.py
├── workflows/          # qwen21_t2i.json / qwen21_edit.json / VNCCS_PoseStudio_QI2.1.json
├── prompts/            # the 3 author prompts + their parameters
├── references/         # showcase images (L0/L1/L2 seaside set) + default character image
├── downloads/samples/  # author originals + author metadata + original workflow JSON
└── pose_library/       # 55 hand-picked poses (copied into ComfyUI at install time)

Notes

• Original code and docs in this repo (scripts/, workflows/, prompts/, SKILL.md, README, …) are licensed under the Apache License 2.0
• Third-party assets are not covered by Apache-2.0 and remain with their respective authors: the author originals in downloads/samples/ (Civitai authors), the community poses and previews in pose_library/ (MIUProject / Totemistyk, of which the Ballet repo is MIT), and the models and LoRAs (HF Comfy-Org/Qwen-Image-2.1, Civitai 2986844 / 2958918 / 2988215, RunningHubAI, MIUProject, …)
• The NSFW sample images are for demonstrating the recipe only; this is a public repository, so if that bothers you, star it first and read later

Acknowledgements

The pose-control capability (Routes ③④) and the 136-pose library build on the open-source work of AHEKOT (VNCCS_Utils), Totemistyk and MIUProject. The base model and LoRAs come from Qwen / Comfy-Org, RunningHubAI, and tawsLoRa, TheseAlpacas and posture-v3 on Civitai. Thanks to all of them.