A pose studio for one very specific whale-girl

What it does

Lanpose (澜姿) is a browser-based image studio: you pick pose sketches, a costume, and optionally a face reference, and it generates images of the same model in that pose and outfit. The default model is Dafeiyu, a community whale-girl mascot — long blue hair, whale ears, whale tail, maid outfit. It works standalone, as a DeepSeek Harness plugin, or via direct connection to a local or remote ComfyUI.

The interesting bit

The heavy lifting is done by careful reference-image ordering: model photo, optional face, optional person-free costume image, then the pose line-art — fed in that order to Seedream, Gemini, or any OpenAI-compatible endpoint. The pose sketch contributes only body posture; the blank body from the line-art never bleeds into the result. Custom costume references are generated as a model wearing the outfit, then the person is removed, leaving just the clothes.

Key highlights

• Draws on three external catalogs by artist ギア: 288 pose groups (1,439 sketches), 1,407 costumes, and 1,499 CC0 face references — images fetched on demand, only IDs and keywords stored in the repo.
• As a Harness plugin, agents get search_poses, search_costumes, search_faces, and generate_pose_images tools; without an API key it still returns ready-to-use prompts for dsh-image-gen.
• Adult poses and costumes are gated behind an explicit include_adult: true — off by default in both UI and agent search.
• Custom poses and costumes (uploaded or AI-generated) stay browser-local and surface in agent searches tagged source: custom.
• Generation history persists across refreshes, with per-item parameter recall and a “reuse” action.

Caveats

• Deeply tied to one character and one artist’s catalogs; importing up to three other model photos is supported, but this is not a general-purpose tool.
• Licensing is a patchwork: the mascot design is CC BY-NC-SA (non-commercial), pose sketches are non-commercial per upstream, and the repo’s MIT covers only the code.
• OpenAI-compatible endpoints that lack image editing silently fall back to text-to-image, with a note that reference images weren’t sent.

Verdict

Worth a look if you build agent tooling and want a well-executed example of reference-ordered image generation, or if you happen to be a Dafeiyu fan. Everyone else can admire the plumbing from a distance.