Arekejoker/qwen21-nsfw-pose-skill

The problem

Qwen-Image 2.1 can already make decent images, but getting consistent, promptable, deliberately posed output in ComfyUI is a matter of trial and error: which sampler, which CFG, what prompt structure, and how big an NSFW LoRA can be turned up before the image falls apart. The knowledge usually lives in someone else's Civitai sample page. This repo packages a working recipe — base model version, sampler/scheduler pair, step count, prompt skeleton, LoRA name and weight — plus a pose-control route and a self-check script, so a run starts from a known-good configuration instead of guesswork.

What's different

Most NSFW ComfyUI workflows circulate as bare JSON files or screenshots in Discord; you get a graph but not the reasoning. Here the emphasis is on a forced recipe with a diagnostic step: doctor.py checks model files, LoRA, custom nodes and browser backend before a run, and the docs state explicit degradation rules (don't quietly weaken the recipe without telling the user). The second differentiator is pose control: instead of relying on prompt words for body position, it routes through a 3D puppet (MakeHuman mesh rendered via three.js, driven by the VNCCS node) so the same character can be re-posed while keeping identity. The 136-pose library is assembled from public HuggingFace collections plus hand-picked items. Compare with a plain Civitai workflow (no checks, no pose geometry) or with ControlNet-style pose transfer (needs a clean reference image; the README itself warns that photos/illustrations get blended into the output, so only identity-free doll renders work).

Who it's for

People who already have ComfyUI running, own a 12GB+ NVIDIA card (24GB recommended) or a large-memory Apple Silicon machine, and want a repeatable NSFW/pose workflow rather than one-off lucky seeds. It also fits agent users (Claude Code, Codex, Cursor, OpenClaw) who want a skill that invokes the scripts. It is not for CPU-only setups, 8GB cards without accepting GGUF quantization and slow offload, or anyone who wants a one-click hosted service — this is local-first and dependency-heavy.

Cost of adoption

About 16GB of base model files, up to ~40GB of disk, plus separate LoRA downloads from Civitai (registration often required). Route ④ needs a desktop browser reachable via Playwright/CDP. macOS numbers in the README are explicitly community-reported and untested by the maintainer; edit mode has a known MPS VAE bug with a --cpu-vae workaround. Licensing is split: the repo code is Apache-2.0, but sample images, pose assets and LoRAs belong to their original authors and carry their own terms. The README is unusually detailed on hardware and failure modes, which helps.

Verdict

The project is 11 days old with ~3.7k stars and roughly +318 stars/day, ranked #19 on a new-repo board but never on the daily trending board — that pattern suggests a fast, topic-driven spike rather than an established tool. The content is substantive (author-replicated samples with metadata, a doctor script, documented trade-offs), which is more than most NSFW recipe repos offer, but an 11-day-old repository has no track record on maintenance or breaking changes in ComfyUI or the base model. Treat it as a good starting configuration to fork and pin, not as infrastructure you depend on unexamined.