Skip to content
Back to search

qwen21-nsfw-pose

Use when generating sexy/NSFW or posed images with local Qwen-Image 2.1 — pose/NSFW LoRAs (tawsLoRa femaledVaginus, TheseAlpacas, posture-v3), pose transfer from an existing image, or exact pose control via VNCCS PoseStudio 3D pose library. Triggers: 性感姿势, 摆姿势, NSFW 出图, 换姿势, 姿势控制, pose transfer, pose lora, VNCCS, 人偶姿势, 姿势库. Independent of any general image-generation skill (which handles SFW text/edit/cutout work).

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/arekejoker--qwen21-nsfw-pose-skill/badge.svg)](/listings/arekejoker--qwen21-nsfw-pose-skill)

HTML

<a href="/listings/arekejoker--qwen21-nsfw-pose-skill"><img src="/api/listings/arekejoker--qwen21-nsfw-pose-skill/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 403 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: qwen21-nsfw-pose
  3. description: Use when generating sexy/NSFW or posed images with local Qwen-Image 2.1 — pose/NSFW LoRAs (tawsLoRa femaledVaginus, TheseAlpacas, posture-v3), pose transfer from an existing image, or exact pose control via VNCCS PoseStudio 3D pose library. Triggers: 性感姿势, 摆姿势, NSFW 出图, 换姿势, 姿势控制, pose transfer, pose lora, VNCCS, 人偶姿势, 姿势库. Independent of any general image-generation skill (which handles SFW text/edit/cutout work).
  4. license: Apache-2.0
  5. ---
  6. # Qwen-Image 2.1 NSFW / Pose Pipeline
  7. Local photorealistic NSFW + pose control: ComfyUI + Qwen-Image 2.1 (int8 convrot)
  8. + NSFW LoRAs + a 136-pose VNCCS 3D pose library.
  9. **Path convention:** everything in this skill is relative to the skill root
  10. (this directory). Scripts in `scripts/`, workflow graphs in `workflows/`,
  11. author prompts in `prompts/`, reference images in `references/`,
  12. author sample data (originals + metadata) in `downloads/samples/`.
  13. ComfyUI-side locations (models, LoRAs, pose library) are configurable via env
  14. vars — see `install/setup.md` and §2. The default host for ComfyUI is
  15. `http://127.0.0.1:8188` (override with `COMFYUI_URL`).
  16. This file is self-contained: setup, recipe, pose selection, and a first-run
  17. author case are all inside.
  18. ## 0. First run — follow the author's case (recommended)
  19. The fastest way to trust the pipeline is re-running the author's own sample
  20. prompt and comparing against the author's published image (verified 2026-10-12):
  21. ```bash
  22. python3 scripts/qwen_pose_test.py \
  23. --mode t2i --name author_case_shower \
  24. --prompt "$(cat prompts/taws_v13_shower.txt)" \
  25. --seed 80964 \
  26. --steps 50 --cfg 2.5 --sampler er_sde --scheduler beta \
  27. --width 1152 --height 1728 \
  28. --lora "qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors" \
  29. --lora-strength 1.0 \
  30. --out outputs/author_case/shower.png
  31. ```
  32. - Other cases: `prompts/taws_v13_massage.txt` (30 steps, same seed, per its
  33. `.params.json`) and `prompts/thesealpacas_v2_selfie.txt` (CFG 4.0 +
  34. `thesealpacas-nsfw-v2` @0.8; its `.txt` ends with a `[negative]` section —
  35. pass that to `--negative`).
  36. - Author originals to compare against: `downloads/samples/*_author.jpg`
  37. (shower / massage / selfie). Full author metadata (prompts/params/seed/
  38. complete ComfyUI graph) in `downloads/samples/phase5_meta.json`.
  39. ## 1. Routing
  40. | Request | Route | Tool |
  41. | --- | --- | --- |
  42. | New sexy/NSFW image from a text idea | ① T2I + NSFW LoRA (mandatory) | `scripts/qwen_pose_test.py --mode t2i` |
  43. | Change pose of an existing character image (keep face/clothes) | ② Edit | `scripts/qwen_pose_test.py --mode edit` |
  44. | Change pose from a 3D mannequin render (script, no browser) | ③ Pose transfer | `scripts/run_pose_transfer.py` |
  45. | Exact pose from the 3D pose library (136 poses) | ④ VNCCS PoseStudio | `scripts/run_vnccs_ui.py` (browser automation) |
  46. | Do NOT use: a plain photo or illustration as pose reference — the model composites it as a second person, or redraws the reference person entirely (tested twice); use a textureless 3D mannequin render with route ③/④ instead | | |
  47. Outputs land under `./outputs/` (the scripts create directories as needed).
  48. ## 2. Setup — what this package contains and what to install
  49. ### 2.1 In this package
  50. | Path | Content |
  51. | --- | --- |
  52. | `SKILL.md` | this file |
  53. | `scripts/` | `doctor.py` (read-only environment self-check: ComfyUI, models, LoRAs, VNCCS, pose library, browser stack — every check degrades to WARN/SKIP with a fix hint), `qwen_pose_test.py` (routes ①②), `run_pose_transfer.py` (route ③: pose transfer from a 3D mannequin render, no browser), `run_vnccs_ui.py` (route ④), `make_custom_pose.py` (new pose from a library pose by bone rotation), `run_phase5.py` + `build_phase5_sheet.py` (case re-run & comparison sheet), `media_lock.py` (shared media lock, bundled), `browser_fallback.py` (standalone CDP fallback when the OpenClaw adapter is absent), `validate_package.py` (offline package integrity check) |
  54. | `workflows/` | `qwen21_t2i.json`, `qwen21_edit.json` (official Qwen-Image 2.1 graphs), `VNCCS_PoseStudio_QI2.1.json` |
  55. | `prompts/` | the 3 author prompts + `.params.json` (the exact local-test settings) |
  56. | `references/` | `pose_examples/` (L0/L1/L2 seaside examples), `src_standing.png` (default character for route ④) |
  57. | `downloads/samples/` | author originals `*_author.jpg` (shower/massage/selfie) + `phase5_meta.json` (prompt, params, seed, complete graph JSONs) + the 3 raw ComfyUI graph files — used by `scripts/run_phase5.py` & `build_phase5_sheet.py` |
  58. | `pose_library/local_user_poses/Uncategorized/` | 55 curated extra pose JSONs + previews — copy into ComfyUI's `PoseLibrary/local_user_poses/Uncategorized/` after installing the node |
  59. | `install/setup.md` | full install guide (models, LoRAs, node, pose repos, commands) |
  60. ### 2.2 ComfyUI-side dependencies (see `install/setup.md` for exact commands)
  61. - **Base models (3 files)** from HF `Comfy-Org/Qwen-Image-2.1`:
  62. `diffusion_models/qwen_image_2.1_int8_convrot.safetensors`,
  63. `text_encoders/qwen3vl_8b_int8_convrot.safetensors`,
  64. `vae/qwen_image_2.1_vae_bf16.safetensors` — into the ComfyUI model dirs.
  65. - **LoRAs** (into ComfyUI `models/loras/`):
  66. | File (size) | Use | Source |
  67. | --- | --- | --- |
  68. | `qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors` (608M) | default: anatomy + pose, sharpest | Civitai model 2986844 — tawsLoRa "Qwen Image 2.1 - FemaledVaginus-NSFW" |
  69. | `qwen2_1-femaledVaginus-nsfw-v1.3-fine.safetensors` (608M) | same family, v1.3 | Civitai model 2986844 |
  70. | `thesealpacas-nsfw-v2.safetensors` (79.7M) | most reveal, "influencer" face drift | Civitai model 2958918 |
  71. | `nsfw-posture-v3.safetensors` (84M) | strongest pose/body emphasis | Civitai model 2988215 |
  72. | `qwen21-nsfw-image-edit.safetensors` (84M) | edit-route undress/pose changes | HF `RunningHubAI/Qwen-Image-2.1-NSFW-Image-Edit` (mirror of oreolove "Nude Edit"; original Civitai page delisted) |
  73. | `VNCCS_QI2_PoseStudioV1.1.safetensors` (159M) | routes ③ and ④ (file name selectable via the `VNCCS_LORA` env var) | HF `MIUProject/VNCCS_PoseStudio_QI2.1` (same repo as the workflow JSON) |
  74. - **VNCCS_Utils custom node**: https://github.com/AHEKOT/ComfyUI_VNCCS_Utils →
  75. clone into `ComfyUI/custom_nodes/`, then restart ComfyUI once.
  76. - **Pose library**: 4 HuggingFace repos (the public catalog as of 2026-10-12):
  77. `Totemistyk/General_Poses_PoseStudio` (50), `Totemistyk/Ballet_PoseStudio`
  78. (3 in manifest; the repo actually ships 58 — register + refresh, then copy
  79. this package's 55 curated extras into `local_user_poses/Uncategorized/`),
  80. `MIUProject/VNCCS_PoseLibrary_Main` (16), `MIUProject/MySuperPoseRepo` (12).
  81. Register: `curl -X POST 127.0.0.1:8188/vnccs/pose_library/repositories/add
  82. -d '{"repo_id":"<owner>/<repo>"}'`. Total after setup: 136.
  83. **A repo refresh deletes local files not in the remote manifest** — extra/
  84. curated poses belong in `PoseLibrary/local_user_poses/<Category>/`
  85. (always listed, never auto-cleaned). Pose list API: `GET /vnccs/pose_library/list`.
  86. - **Route ④ extras**: any CDP Chrome (default `127.0.0.1:18800`) +
  87. `browser_client.py` + playwright. On the original host these live in
  88. `~/.openclaw/tools/browser/`; elsewhere point `run_vnccs_ui.py` at an
  89. equivalent CDP endpoint (it only needs `connect_sync(playwright)` to return a
  90. browser with one context — a minimal stand-in is fine).
  91. - **media_lock**: bundled in `scripts/`; both scripts import it first from the
  92. package, then fall back to `~/.openclaw/tools`.
  93. The base model is `qwen_image_2.1_int8_convrot.safetensors` — the same model
  94. the Civitai authors of these LoRAs used, so their published prompts/params
  95. transfer well. (Note: one of the three author samples actually ran on a
  96. different base with 5 stacked LoRAs; we reproduced it with the single
  97. thesealpacas-v2 LoRA on the convrot base — see Parameter provenance below.)
  98. Download tips (verified): Civitai via the official CLI or
  99. `curl -H "Authorization: Bearer <token>" <file-url>`; HuggingFace via
  100. `hf-mirror.com` directly (no proxy — a proxy causes SSL_ERROR).
  101. ## 3. Route ① — T2I (the "author recipe", validated 2026-10-12)
  102. ### 3.1 Mandatory recipe — do not weaken silently
  103. This recipe reproduces Civitai showcase-level results. **All of the following are
  104. part of the recipe; a deviation changes the result and must be told to the user:**
  105. - **Sampler** `er_sde` + scheduler `beta` (not euler/simple)
  106. - **CFG 2.5** (the A1 reference; 2.5–4 works, CFG 1 starves the LoRA)
  107. - **Steps = the reference sample's value: 50** (A1 provenance, table below)
  108. - **Size 1152×1728 portrait 2:3** (`--width 1152 --height 1728`); keep
  109. `--resolution 1024` (that flag feeds `TextEncodeQwenImage21`, it is not pixel size)
  110. - **A LoRA is always on for route ①.** Default:
  111. `qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors` @ **1.0**
  112. - **The prompt is the main lever** — it must follow the author structure
  113. (§3.2) at the author's density (300–1500 characters, explicit terms).
  114. Short/keyword prompts and SFW-generic phrasing are what make output look bland.
  115. - **Pass the reference sample's seed** when you claim "reproduction"
  116. (A1 seed = 80964). The harness default seed (424242) is fine for new
  117. creative runs but must not be labelled as the author's result.
  118. Parameter provenance (source: `scripts/run_phase5.py` + each run's `*.png.json`):
  119. | Run record | Steps | CFG | Seed | LoRA strength (local test) |
  120. | --- | --- | --- | --- | --- |
  121. | A1 shower (primary reference) | 50 | 2.5 | 80964 | 1.0 |
  122. | A2 massage (separate sample) | 30 | 2.5 | 80964 | 1.0 |
  123. | B1 selfie | 50 | 4.0 | 124880844757561 | 0.8 (thesealpacas-v2) |
  124. | C1 control | 50 | 2.5 | 80964 | no LoRA |
  125. Preserve each sample's settings as a unit; do not mix one sample's prompt with
  126. another sample's params and present it as "the author recipe".
  127. ### 3.2 Prompt structure (copy the pattern, fill the scene)
  128. The working prompts all contain, in this order — write all 7, no omissions:
  129. 1. Camera: "A hyper-realistic photograph / low-angle shot / top-down overhead shot", lens and angle
  130. 2. Subject: age, build, hair (color, wet/dry, placement)
  131. 3. Pose: pick one from the pose ladder (§4) and spell out each limb (which leg bent, where the hands are, what the head/gaze does)
  132. 4. Setting + lighting: "warm steamy light", "neon bokeh", "soft morning light through sheer curtains"
  133. 5. Skin: "glistening", "visible pores", "natural skin texture", "slick oily sheen"
  134. 6. Anatomy/reveal where desired — the LoRAs were trained on explicit terms
  135. (e.g. "labia majora ... innie distinctly visible"); generic words ("seductive")
  136. do almost nothing
  137. 7. Finish: "8k resolution, cinematic lighting, candid 'caught in the moment'
  138. atmosphere, photorealistic anatomy"
  139. Template (a real author prompt runs 300–1500 chars):
  140. ```
  141. A hyper-realistic, {camera} shot {subject with age/build/hair state},
  142. {one pose from §4, limb by limb}. {Setting + lighting}. Her skin is
  143. {glistening/visible pores/slick oily sheen}. {explicit anatomy/reveal wording}.
  144. {face/expression}. 8k resolution, cinematic lighting, candid "caught in the
  145. moment" atmosphere, photorealistic anatomy.
  146. ```
  147. **Before writing, load a verified author prompt and rewrite it for your scene** —
  148. `prompts/*.txt` (+ matching `.params.json`). Writing from scratch is what makes
  149. results look generic.
  150. ### 3.3 Deviations — tell, don't silently do it
  151. If the user asks for a shorter prompt, a SFW-generic prompt, "no LoRA", fewer
  152. steps, or a different sampler — **and** if you (the agent) ever need to deviate
  153. from the recipe for any reason (model refuses a token, out-of-memory, etc.):
  154. say the expected cost out loud, show the evidence, and proceed only after
  155. confirmation. Never present a weakened run as "the recipe's result"; label what
  156. was changed next to the delivered image. Evidence to cite:
  157. - **True controlled experiment (single variable = LoRA):** `C1` (author prompt,
  158. 50 steps, seed 80964, **no LoRA**) vs `A1` (identical prompt/params + LoRA):
  159. no-LoRA still looks showcase-grade as a photograph, but the LoRA is what adds
  160. the skin/gloss/anatomy detail the author's images have. So "no LoRA" ≈
  161. "good photograph, not the author's texture".
  162. - **Visual demo (multiple variables: prompt + outfit + LoRA + seed):**
  163. `references/pose_examples/01_seaside.png` (L0, no LoRA) vs
  164. `01b_seaside_lora.png` (L1 + LoRA) vs `06_l2_squat_lowangle.png`
  165. (L2 + low angle + LoRA) — shows the ladder of how much the full recipe adds.
  166. It demonstrates the effect but does **not** isolate any single factor.
  167. ### 3.4 Known limits of the harness
  168. - **One LoRA per run.** `qwen_pose_test.py` injects a single `LoraLoaderModelOnly`
  169. (node id 99) between `UNETLoader` and `KSampler`. Civitai showcase workflows
  170. often stack 3–5 LoRAs; to reproduce those, chain extra `LoraLoaderModelOnly`
  171. nodes in `build_workflow()`.
  172. - **No refine pass / no EasyCache.** Authors add a 2× lanczos upscale + a 25-step
  173. `denoise 0.35` pass and `EasyCache`; `QwenImage21Cache` is broken on the
  174. original host (`aimdo memory compile error`) and the route ④ runner removes it.
  175. A/B runs matched the authors' look without the refine pass — add it only when
  176. a job needs extra sharpness.
  177. ## 4. Pose selection (the "boldness scale") — match the user's request
  178. Read the user's request (clothes state, desired mood, scene) → pick a tier →
  179. give a **concrete pose**, not a vague one. "Bold" is three levers combined:
  180. **pose openness × camera angle × clothing/wet state**.
  181. ### 4.1 Tiers (English wording to use in route ① prompts)
  182. - **L0 closed / editorial** (clothed SFW): standing with weight on one leg, hand
  183. on hip or in pocket; sitting on a chair, legs crossed at the ankles; leaning on
  184. a railing/wall, looking back over the shoulder; mid-stride walking.
  185. - **L1 open** (swimwear / wet / relaxed): crouching with knees apart, hands on
  186. the floor; sitting on the floor with knees opened to one side; lying on her
  187. side in bed, propped on an elbow, looking back over the shoulder; kneeling with
  188. hands behind; reclining on a sofa with legs open.
  189. - **L2 bold** (author style — always pair with the NSFW LoRA + low angle):
  190. standing with legs apart, shot from a low angle looking up; lying prone with
  191. hips lifted toward the camera; sitting with legs wide open, leaning forward on
  192. the elbows; kneeling on one knee with the other leg extended to the side, torso
  193. arched back; lying on her back with legs open, one hand behind the head.
  194. - **L3 extreme / cropped**: extreme low angle looking up between open legs;
  195. overhead shot lying flat with legs open; cropped framing (waist-up or
  196. mid-thigh) to emphasize the pose.
  197. ### 4.2 Suggestion table (user request → tier → concrete pose)
  198. | User asks for (examples) | Tier | Route ④ library pose families (`--list-poses` names) | Route ① prompt wording |
  199. | --- | --- | --- | --- |
  200. | 全着装、自然/编辑感 (standing, portrait, office/street) | L0 | Standing / Sitting / Walking / Close Up | standing with weight on one leg, hand in pocket; seated on a chair, legs crossed at the ankles |
  201. | 泳装/湿身/海滩/浴室 (swimwear, wet, relaxed) | L1 | Squatting / Kneeling / Laying Back | crouching with knees apart, hands on the floor, looking back over the shoulder |
  202. | 大胆/NSFW (nude + LoRA, low angle, "author style") | L2 | Kneeling / Laying Back / Prone / Splits / Crawling / Dancing on knees | standing with legs apart from a low angle; prone with hips lifted toward the camera; seated wide, leaning forward on elbows |
  203. | 极端/特写构图 (crop, extreme low, overhead) | L3 | Splits / any L2 family + crop framing | extreme low angle looking up between open legs; overhead, lying flat with legs open |
  204. Rules:
  205. 1. Decide the reveal first: clothed → L0; swimwear/wet → L1; nude + NSFW LoRA → L2.
  206. **Pairing the LoRA with an L0 pose is the most common cause of "bland" NSFW
  207. output** (observed, not a proven law).
  208. 2. Camera angle amplifies the pose (empirical heuristic): low angle ≈ +1
  209. boldness, overhead ≈ +1. "Low angle + L1" ≈ "eye level + L2" — this is why
  210. the author's shower/massage shots read as bold despite simple poses.
  211. 3. One pose per prompt, limb by limb; two competing pose descriptions get blended.
  212. 4. When the user names a pose they saw somewhere, search the library first:
  213. `python3 scripts/run_vnccs_ui.py --list-poses` (136 entries with file paths).
  214. ## 5. Route ② — Edit (pose change with identity preservation)
  215. ```bash
  216. python3 scripts/qwen_pose_test.py \
  217. --mode edit --name pose_change --input /absolute/base.png \
  218. --prompt "change her pose to kneeling on one knee, hands on hips, keep face hair and outfit" \
  219. --cfg 3 --sampler er_sde --scheduler beta --steps 20 \
  220. --out outputs/pose_change.png
  221. ```
  222. CFG 3 is more stable for edit mode. SFW instruction prompts keep NSFW LoRAs from
  223. triggering; add `--lora qwen21-nsfw-image-edit.safetensors` only when undressing.
  224. ## 6. Route ③ — Scripted pose transfer (3D mannequin reference)
  225. ```bash
  226. python3 scripts/run_pose_transfer.py \
  227. --char /absolute/character.png \
  228. --pose-ref /absolute/doll_render.png \
  229. --out outputs/pose_transfer.png [--seed N] [--steps 25]
  230. ```
  231. The pose source must be a textureless **3D mannequin render** — e.g. the
  232. `doll_*.png` captures that route ④ saves under the ComfyUI output dir, or any
  233. 3D poser render of the target pose. Photos and illustrations as pose references
  234. do NOT work: the model composites the reference person as a second person, or
  235. redraws the reference person entirely (retested 2026-10-16 — all variants
  236. failed).
  237. Verified recipe (2026-10-16) — the script sets it up automatically:
  238. - dual-image edit with **image 1 = mannequin** (pose) and **image 2 = character** (identity);
  239. - prompt `replace the pose of <image 2> with the pose of <image 1>. keep the character of <image 2>`;
  240. - **VNCCS PoseStudio LoRA @ 1.0 is mandatory** — without it the mannequin is
  241. ignored entirely (3 no-LoRA controls);
  242. - 25 steps, cfg 1, euler/simple, 1024 px. Single person, identity kept, pose
  243. follows the render; stable across seeds (sit ×3 seeds + crawling /
  244. kneel-side / lean-back all passed). Camera framing can differ slightly from
  245. the render; producing *new* mannequin renders still requires a route ④
  246. browser session — route ③ reuses captures you already have.
  247. ## 7. Route ④ — VNCCS PoseStudio (exact pose from the 3D library)
  248. The `VNCCS_PoseStudio` node needs the ComfyUI **web frontend** for its 3D "live
  249. capture"; the headless API fails with "Live capture timed out". So the runner
  250. drives a CDP browser: load the official graph → set character image + pose
  251. JSON → queue → collect outputs from disk. ~60–90 s each.
  252. ```bash
  253. # list all 136 poses with resolved file paths
  254. python3 scripts/run_vnccs_ui.py --list-poses
  255. # run one pose on a character image (+ optional scene prompt)
  256. python3 scripts/run_vnccs_ui.py --pose "squatting leaning back" \
  257. [--char /absolute/face.png] [--prompt "on a rooftop at night, city bokeh"] [--timeout 900]
  258. ```
  259. - `--pose` takes the name/slug case-insensitively (`crawling`, `Squatting
  260. Leaning Back`, `arabesque`). 9 names exist in more than one repository
  261. (`Standing A`, `CloseUp A`, …); qualify with `<repo>/<category>/<name>`
  262. (exact keys printed by `--list-poses`, e.g.
  263. `miuproject_vnccs_poselibrary_main/standing/standing_a`).
  264. - **Scene prompt (`--prompt`):** the capture prompt is rebuilt at execution
  265. time from the scene document's `export` block (default template
  266. `Draw character from image2\n<lighting>\n<user_prompt>`); `--prompt "..."`
  267. injects `export.user_prompt` before the graph loads. Editing the
  268. `TextEncodeQwenImage21` widget or adding `export.user_prompt` to a *library
  269. pose file* has no effect. To change the scene of an existing route ④ result,
  270. re-run it through route ② (`qwen_pose_test.py --mode edit`).
  271. - The runner reaches a CDP browser through the OpenClaw adapter when present,
  272. otherwise through the bundled `scripts/browser_fallback.py` (reuses a
  273. reachable `COMFY_BROWSER_CDP` endpoint, else launches a throwaway headless
  274. Chromium; `COMFY_BROWSER_MODE=auto|attach|launch`).
  275. - The runner re-indexes the pose library on every call — newly downloaded poses
  276. are usable immediately. Each run writes `<pose>_<stamp>.meta.json` next to
  277. the outputs and never clobbers previous results (unique comfy prefix per run).
  278. - Output: `./outputs/vnccs_ui/` (override with `VNCCS_UI_OUT`). The 3D doll is a
  279. bare MakeHuman mannequin — nude prompts + pose transfer can produce nude
  280. output (expected: the model reads the doll's state).
  281. - **Custom poses from code:** `scripts/make_custom_pose.py` loads a library
  282. pose into the PoseStudio editor, adds bone rotations, and saves through the
  283. editor's own Save (JSON + preview in
  284. `PoseLibrary/local_user_poses/poses/<category>/`). The new pose works with
  285. `run_vnccs_ui.py --pose` at once. `--list-bones` prints the base pose's
  286. euler angles; pass the full library id when a name is not unique.
  287. ```bash
  288. python3 scripts/make_custom_pose.py \
  289. --base "MIUProject__VNCCS_PoseLibrary_Main/poses/Standing/Standing A" \
  290. --name "Arms Out" --rotate upperarm_l:1:90 --rotate upperarm_r:1:-90
  291. ```
  292. - **Pose fidelity caveat:** each pose JSON carries its own `cameraParams`
  293. (yaw/pitch/zoom); the generated image inherits that viewpoint. Poses whose
  294. shape only reads from another angle (e.g. a true arabesque's extended back
  295. leg) come out front-facing and ambiguous — rotate the doll in the UI and
  296. re-run, or pick a pose that reads from the front.
  297. ## 8. Environment & Safety
  298. - ComfyUI as a service (original host: `systemctl --user status|restart
  299. comfyui.service`). After installing new custom nodes, one restart is required.
  300. - The route scripts take the bundled media lock in shared mode
  301. (`media_lock(exclusive=False, wait_timeout=...)`) so they don't fight video
  302. jobs on a shared machine; route ④ wraps the whole browser run in it.
  303. - On the original host, video tasks hold the lock exclusively (~06:45 / 09:00 /
  304. 11:30 openclaw windows); expect waits, not conflicts.
  305. - If a graph includes `QwenImage21Cache` and the host errors on it
  306. (`aimdo memory compile error`), remove the node and rewire LoRA → KSampler
  307. (the route ④ runner does this automatically in the browser).
  308. - Never modify the workflow JSONs in this package's `workflows/` while a run is
  309. in flight; the harness works on in-memory copies anyway.
  310. - A 50-step 1152×1728 run takes ~200–400 s on the int8 model; a 30-step edit
  311. run ~60–120 s.
  312. ## 9. Verification
  313. 1. Check the harness printed `ok: true` and the output file is non-empty.
  314. 2. Visually inspect with an image-reading tool (composition, pose match,
  315. anatomy, face). For pose-transfer runs, compare against the reference.
  316. 3. For reproductions of published images, compare side-by-side with the author
  317. original (`downloads/samples/*_author.jpg`; comparison-sheet
  318. example: `scripts/build_phase5_sheet.py`).
  319. 4. Record prompt/seed/steps/LoRA in the run's `*.json` meta file (harnesses
  320. write one automatically for routes ①②③④).
  321. 5. When a recipe deviation was made (weaker prompt, no LoRA, fewer steps,
  322. default seed), label the output and state the deviation next to it in the
  323. delivery message.
  324. Last full verification: 2026-10-12 — route ① (4 author-prompt reproductions),
  325. route ② (9/9 pose edits with identity preserved), route ④ (6 pose runs:
  326. crawling, kneeling side, leaning back, sitting side, squatting leaning back,
  327. arabesque — the last one verified for output only, pose-shape match unverified;
  328. its doll capture turned out to be a stale legacy capture),
  329. pose index count 136 == `GET /vnccs/pose_library/list`, pose-ladder A/B images
  330. in `references/pose_examples/`.
  331. Follow-up checks 2026-10-14: route ④ `--prompt` (scene follows the prompt, pose
  332. preserved), bundled browser fallback launch path, and custom pose creation via
  333. `scripts/make_custom_pose.py` (save + output verified).
  334. Follow-up checks 2026-10-16 (route ③): 6 mannequin-reference transfers passed
  335. (sit ×3 seeds, crawling, kneel-side, lean-back): single person, identity kept,
  336. pose follows the render; 3 no-LoRA controls ignored the mannequin; photo and
  337. illustration references still fail (composite / redraw of the reference
  338. person). Script `scripts/run_pose_transfer.py` added and end-to-end tested.
README.md · README · 148 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # qwen21-nsfw-pose
  2. [![Package checks](https://github.com/Arekejoker/qwen21-nsfw-pose-skill/actions/workflows/validate.yml/badge.svg?branch=main)](https://github.com/Arekejoker/qwen21-nsfw-pose-skill/actions/workflows/validate.yml)
  3. [English](README.en.md) | 中文
  4. **Qwen-Image 2.1 本地 NSFW / 姿势出图完整配方** — 作者级提示词结构 + NSFW LoRA + 136 个 3D 姿势库 + VNCCS 人偶精确姿势控制。
  5. 基于 ComfyUI + `qwen_image_2.1_int8_convrot` 底模(与 Civitai 作者同款底模),复刻并验证过 Civitai 作者样图(2026-10-12)。
  6. - 适用:Claude Code / Codex / Cursor / OpenClaw(agent 技能)+ 人类手动使用
  7. - 内容:可跟跑的作者案例、强制配方、姿势大胆度阶梯(L0–L3)、完整依赖与安装指南
  8. **演示**(3D 人偶调姿势 → 一键生成 → 成品,Route ④):
  9. <img src="references/demo_pose_studio.gif" width="100%" alt="演示:3D 人偶变形 → 实时生成 → 成品" />
  10. ---
  11. ## 效果展示
  12. ### 本配方输出:同一海边场景的三级对照(L0 → L1 → L2)
  13. | L0 · 全着装,无 LoRA(提示词即全部) | L1 · 湿身泳装 + LoRA@1.0 | L2 · 极低机位 + 湿身蹲姿 + LoRA@1.0 |
  14. | --- | --- | --- |
  15. | <img src="references/pose_examples/01_seaside.png" width="290" /> | <img src="references/pose_examples/01b_seaside_lora.png" width="290" /> | <img src="references/pose_examples/06_l2_squat_lowangle.png" width="290" /> |
  16. 三张图同场景递进:L0 只有提示词;L1 加 LoRA + 显式皮肤/水珠描写;L2 再加低机位与开合姿态。**LoRA 负责质感,提示词负责信息密度,机位负责大胆度。**
  17. ### NSFW 效果(⚠️ 以下图片包含裸露内容)
  18. <details>
  19. <summary><picture><img src="references/author_samples_cover_cursor.gif?v=3" width="100%" alt="作者原图(NSFW)· 点击卡片展开" /></picture><br/><b>👇 点击上方卡片展开作者原图(含裸露内容 · 18+,再点收起)</b></summary>
  20. **作者原图**(Civitai showcase,本配方复刻的目标):
  21. | 作者 · 淋浴(tawsLoRa) | 作者 · 按摩(tawsLoRa) | 作者 · 自拍(TheseAlpacas) |
  22. | --- | --- | --- |
  23. | <img src="downloads/samples/taws_v13_shower_author.jpg" width="300" /> | <img src="downloads/samples/taws_v13_massage_author.jpg" width="300" /> | <img src="downloads/samples/thesealpacas_v2_selfie_author.jpg" width="300" /> |
  24. </details>
  25. 用 §「快速开始」的命令可复刻淋浴那张(复刻输出已验证与作者原图同场景同构图,见 `downloads/samples/phase5_meta.json` 与本文快速开始一节)。
  26. ### 3D 人偶精确姿势(Route ④,136 姿势库)
  27. 默认人物图 `references/src_standing.png`;姿势 = 骨骼旋转 JSON(MakeHuman 网格 + three.js 渲染),
  28. 换姿势且保身份。姿势库来源:HuggingFace 4 个公开仓库(`Totemistyk/General_Poses`、
  29. `Totemistyk/Ballet_PoseStudio`、`MIUProject/VNCCS_PoseLibrary_Main`、
  30. `MIUProject/MySuperPoseRepo`)+ 55 个手工精选,共 **136 个**。
  31. ---
  32. ## 快速开始(作者案例跟跑)
  33. 前置安装见 [install/setup.md](install/setup.md)。装好后先跑一次自检(只读,不改动任何东西):
  34. ```bash
  35. python3 scripts/doctor.py
  36. ```
  37. 它逐项检查 ComfyUI 连通性、底模/LoRA、VNCCS 节点、姿势库、浏览器后端,并给出缺失项的修复指引。
  38. 全绿后一条命令复刻作者淋浴样图(~5 分钟):
  39. ```bash
  40. python3 scripts/qwen_pose_test.py \
  41. --mode t2i --name author_case_shower \
  42. --prompt "$(cat prompts/taws_v13_shower.txt)" \
  43. --seed 80964 \
  44. --steps 50 --cfg 2.5 --sampler er_sde --scheduler beta \
  45. --width 1152 --height 1728 \
  46. --lora "qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors" \
  47. --lora-strength 1.0 \
  48. --out outputs/author_case/shower.png
  49. ```
  50. 对比对象:`downloads/samples/taws_v13_shower_author.jpg`。
  51. 完整元数据(提示词/参数/seed/完整工作流):`downloads/samples/phase5_meta.json`。
  52. ### 硬件要求(参考)
  53. 底模为 Qwen-Image 2.1 int8 套件,模型文件合计约 16GB(diffusion 6.8GB / text encoder 8.7GB / VAE 0.6GB),LoRA 各自 ≤608MB。
  54. - **最低**:NVIDIA 12GB 显存 + 32GB 内存——能跑,ComfyUI 会自动把部分模型卸载到内存,出图慢一些
  55. - **推荐**:24GB 显存(3090/4090 级)+ 32–64GB 内存——模型基本常驻,体验最顺
  56. - **8GB 显存**:建议用 GGUF 量化版(社区实测 12–16GB 显存可跑 Q4;8GB 靠 offload 会明显更慢)
  57. - 纯 CPU 不推荐(单张图耗时过长)
  58. - 硬盘预留 ~40GB;CPU 无特殊要求(现代 x86_64/ARM 均可);Linux 最省事;Windows 可原生运行基础脚本(doctor / `--list-poses` / 媒体锁回退,社区贡献),完整流程仍建议 WSL2;Mac(Apple Silicon)见下方单独小节
  59. - 路线④额外需要一个可连接的桌面浏览器(Playwright/CDP,见 `install/setup.md`)
  60. 本仓库案例图与实测数据来自统一内存设备(DGX Spark):1024×1024、40 步约 32 秒/张,峰值增量约 7–10GB。
  61. ### macOS(Apple Silicon)
  62. - 适用 M 系芯片(M1 及以上);Intel Mac 不建议
  63. - **模型选择看统一内存**:≥32GB 直接用本仓库原版 int8 套件(ComfyUI + MPS);16–32GB 建议改用 GGUF 量化版(搭配 ComfyUI-GGUF;社区实测 Q8 更快、Q4 更省内存)
  64. - **速度预期:分钟级/张**——社区报告:M5 Pro 48GB 跑 int8 约 144 秒/1024;M1 Max 64GB 约 7–10 分钟/张(加缓存优化节点如 Comfyui-Spectrum-Qwen2.1 可提速 2–3 倍)。对比本仓库 DGX Spark 实测 32 秒/张,仅供参考
  65. - 已知坑:编辑模式在 MPS 上偶发 VAE 编码异常,可用 ComfyUI 的 `--cpu-vae` 绕过(ComfyUI issue #16433,已有修复讨论);内存紧张时建议单张生成、关闭其他大内存应用
  66. - 路线兼容:①②③ 脚本在 macOS 可直接运行(纯 Python 标准库、POSIX 兼容);④ 需 `playwright install chromium`(macOS 版),`install/setup.md` 中的 AppArmor 提示不适用
  67. - ⚠️ 以上 Mac 数据来自社区报告,维护者没有 Mac 硬件、未经实测;如果你在 Mac 上跑通或遇到问题,欢迎开 issue 反馈,我们会持续更新本说明
  68. ## 四条路线
  69. | 需求 | 路线 | 命令 |
  70. | --- | --- | --- |
  71. | 新 NSFW/性感图(文字出图) | ① T2I + NSFW LoRA(必开) | `scripts/qwen_pose_test.py --mode t2i` |
  72. | 已有角色图换姿势(保脸/保服装) | ② Edit | `scripts/qwen_pose_test.py --mode edit` |
  73. | 已有 3D 人偶姿势渲染图,脚本换姿势(不开浏览器) | ③ 脚本姿势迁移(人偶参考 + VNCCS LoRA) | `scripts/run_pose_transfer.py --char <角色图> --pose-ref <人偶渲染图> --out <出图.png>` |
  74. | 指定 3D 姿势库里的精确姿势 | ④ VNCCS PoseStudio | `scripts/run_vnccs_ui.py --pose <name> [--prompt "<场景>"]` |
  75. 路线③的人偶参考图可由路线④的运行自动保存(输出目录里的 `doll_*.png`),或任何 3D 姿态渲染图。
  76. 不要用照片/插画当姿势参考——模型会把参考人物拼进画面,或直接重画参考人物;请使用无身份特征的 3D 人偶渲染图(路线③/④)。
  77. 姿势怎么选(大胆度阶梯 L0–L3、用户要求→姿势族映射表):见 `SKILL.md` §4,或
  78. `scripts/run_vnccs_ui.py --list-poses` 列出全部 136 个姿势。
  79. ## 配方要点(摘要)
  80. - `er_sde` + `beta`,CFG 2.5,50 步,1152×1728 竖版
  81. - **NSFW LoRA 永远开**(默认 `femaledVaginus v1.3b` @ 1.0)
  82. - 提示词必须按作者结构写满 7 要素、300–1500 字符、含 LoRA 触发词
  83. - 弱化配方(无 LoRA/短提示词/减步数)前必须告知用户代价,不许默默减弱
  84. 完整规则与证据:`SKILL.md`。
  85. ## 目录结构
  86. ```
  87. ├── SKILL.md # agent 技能(完整规则)
  88. ├── CHANGELOG.md # 版本变更记录
  89. ├── install/setup.md # 安装指南(模型/LoRA/节点/姿势库)
  90. ├── scripts/ # doctor.py 自检、出图/自定义姿势脚本、浏览器 fallback、media_lock.py、validate_package.py
  91. ├── workflows/ # qwen21_t2i.json / qwen21_edit.json / VNCCS_PoseStudio_QI2.1.json
  92. ├── prompts/ # 3 篇作者原提示词 + 参数
  93. ├── references/ # 效果展示图(L0/L1/L2 海边对照)+ 默认人物图
  94. ├── downloads/samples/ # 作者原图 + 作者元数据 + 原始工作流 JSON
  95. └── pose_library/ # 55 个手工精选姿势(安装时拷入 ComfyUI)
  96. ```
  97. ## 说明
  98. - 本仓库原创代码与文档(`scripts/`、`workflows/`、`prompts/`、`SKILL.md`、README 等)采用 [Apache License 2.0](LICENSE)
  99. - 第三方素材不在 Apache-2.0 覆盖范围内,权利归各自原作者:`downloads/samples/` 作者原图(Civitai 作者)、`pose_library/` 社区姿势与预览(MIUProject / Totemistyk,其中 Ballet 仓库为 MIT)、模型与 LoRA(HF `Comfy-Org/Qwen-Image-2.1`、Civitai 2986844 / 2958918 / 2988215、RunningHubAI、MIUProject 等)
  100. - NSFW 示例图仅供配方演示;仓库为公开仓库,介意图片的先收藏后阅读
  101. ## 致谢
  102. 本仓库的姿势控制能力(路线③④)与 136 姿势库基于 [AHEKOT](https://github.com/AHEKOT/ComfyUI_VNCCS_Utils)(VNCCS_Utils)、[Totemistyk](https://huggingface.co/Totemistyk)、[MIUProject](https://huggingface.co/MIUProject) 的开源工作;底模与 LoRA 来自 Qwen / Comfy-Org、RunningHubAI 及 Civitai 的 tawsLoRa、TheseAlpacas、posture-v3。感谢所有作者。

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.