The avatar studio

The agent's face is the bundled dragon until you ask for another. On the phone since 0.1.20 and, from 0.1.23, on the web app and the desktop through the runtime (nanomuse/avatar/studio.py), the studio draws one from a description, in the chat:

1. Ask. In the chat: 「换个形象：一只戴圆眼镜的橘猫」, "new avatar: a robot owl", "change your avatar to a small fox with a scarf". Or on the studio screen — the pen on the avatar in the profile sheet → Avatar studio…, or the studio tile in Settings — the phone's page: the face at the top with what it is, a description, the seven style chips, Draw four, and the rest of this list without a card in the chat (the progress comes over the socket as {"kind": "studio", "current"}). The agent is not run for it.
2. The cost first. A card says what it will take: with the account's model (nanoMuse Cloud), the relay's figure (GET /v1/estimate) next to what is left in your allowance — about ¥1.5 for the eight pictures at qwen-image-3.0's price; with a key of your own, the count, at your provider's prices. Draw or Not now.
3. Four to choose from. The same subject four times — classic colouring, lighter, darker with a small accessory, a playful take — Muse's vinyl-toy style, full body on white. Tap one, or say which (「第二个」, "the first one", "top right"); Redraw for four more.
4. The poses. The pick is the idle still; four edits of it are the other moods — working with headphones at a laptop, waiting with a crystal ball, happy hugging a star, sorry with a sweat drop — the set the dragon has. They land in the workspace under avatar/<face>/<mood>.webp; the profile's avatar becomes the face id; every app on the account shows it (through /api/files/avatar/<face>/<mood>.webp). A pose that could not be drawn shows the idle picture instead.

Where the pictures come from: the chat model's host and key. The relay speaks the OpenAI images API (/v1/images/generations, /v1/images/edits), as does any OpenAI-compatible provider that draws; Alibaba Cloud Model Studio's own host has neither and is called on its native multimodal endpoint with the same key (qwen-image-3.0 by default). Which model draws is [llm] image_model in config.toml or Connections → Image & video models in the app; empty means the relay's image model for the account, qwen-image-3.0 on Model Studio, none elsewhere — and then the chat says so plainly instead of trying.

Once the stills are on, the studio makes the four short clips the phone's does — idle, working, waiting, happy; the same fixed motions — from them, when the endpoint has a video model: the relay's (wan2.2-i2v-flash) or Wan on Model Studio, through the asynchronous video API (an upload, a task, polling, the MP4), [llm] video_model overriding the choice and [llm] video_base_url naming the host of that API when it is not the chat model's (a relaying host such as the showcase gateway sets it for its runtimes). They land in avatar/<face>/<mood>.mp4; the card shows the stage animating while they come, the cost card counts them, and a clip that fails leaves its still. An OpenAI-compatible provider without a video API gives stills only. The web and the desktop play the clips at 44 px and above — the dragon's own four ship with the web app — and show stills below that, in lists and pickers, under prefers-reduced-motion, or when a clip is missing.

GET /api/avatar says whether a face can be drawn here (and with which clip model), which session is under way, and which face the profile wears and what it is — its description, style and model, kept in avatar/<face>/face.json when the poses land; POST /api/avatar/begin|start|choose|cancel drive a session (the chat card and the studio screen use them), POST /api/avatar/moods draws the poses and clips of the worn face again. Candidates of finished sessions are cleared after a day; faces stay in the workspace as long as the profile — or you — want them.

With a nanoMuse Cloud account the face follows you to every device (docs/cloud.md): the stills go to the relay once, the other devices wear them from avatar/sync-<hash>/ with the description and style in their face.json, and a device keeps the face it drew — clips and all — when the account's pictures are the ones it already wears.