dsh-novelai

dsh-novelai

NovelAI image generation inside DeepSeek Harness — three agent tools, a full studio page, and a prompt knowledge base distilled from four Chinese NovelAI tutorials.

Bring your own NovelAI API token. Nothing is sent anywhere except image.novelai.net.

What you get

nai_image	Generate an image from a prompt. The PNG is saved and returned as an image attachment, so the model actually sees what it drew. Supports style presets (厚涂/平涂/水彩/油画/素描/线稿/像素/3D/Q版/漫画/立绘), artist strings ([[[artist:name]]]), and enhance (auto weight notation per model, quality tags, default negative, token budget warnings).
nai_status	Subscription tier, expiry, usage.
nai_prompt	build a prompt from subject/style/artist/scene, check an existing one (weight notation, brackets, V4.5 vs V5 syntax, token budget), or guide for the cheat sheet.
Studio page	http://127.0.0.1:<port>/dsh-novelai/ — parameters, live draft previews, history, download.
Standalone studio	node studio.mjs → http://127.0.0.1:4173/, no DSH required.

Install

dsh plugin --profile web add dsh-novelai
# restart DSH, then open http://127.0.0.1:3080/dsh-novelai/

Or manually: add "dsh-novelai": "^1.0.0" to your profile's package.json dependencies and to dsh.profile.bundles, then pnpm install.

Configure

Click 设置 on the studio page and paste your NovelAI token (starts with pst-), or write ~/.dsh/novelai.json yourself:

{
  "apiKey": "pst-...",
  "proxy": "",
  "outDir": "",
  "model": "nai-diffusion-4-5-full",
  "width": 832,
  "height": 1216,
  "steps": 28,
  "scale": 5,
  "sampler": "k_euler_ancestral"
}

• proxy — leave empty for a direct connection. If you are behind the GFW, point it at your local HTTP proxy, e.g. http://127.0.0.1:7890. Environment overrides: NAI_API_KEY, NAI_PROXY.
• outDir — empty means ~/nai-output.
• Tokens are stored in plain JSON under your home directory. That is the same place DSH keeps its own config; do not commit it.

Anlas: what is free

Opus is unlimited for the free tier only: one image, ≤28 steps, up to about 1 MP.

• 832×1216 at 28 steps, 1 image → free
• 30 steps, or 1216×1216, or n=4 → charged, and a 402 means your balance cannot cover it
• V5 at standard size → also free
• Character Reference / Vibe Transfer → charged separately (reference import + per generation)

You can read your balance for free: send a request that is certain to exceed it (e.g. 1216×1216); the 402 body reports Required: X, Available: Y.

Notes

• One generation at a time per account; a second concurrent request returns 429 for ~15 s.
• V4/V4.5 models require v4_prompt / v4_negative_prompt; this plugin builds them for you.
• The bundled knowledge base (docs/prompt-guide.md) covers style reproduction, V5 specifics, character reference, Anlas accounting, and the precise-reference API shape.
• Reference images (Character Reference / Vibe Transfer) are not implemented yet.

License

MIT