Skip to content

The avatar studio ​

The agent's face is the bundled dragon until you ask for another; inspiration cards do not change that default. 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. Your image provider. Add an image provider and API key under Models first. nanoMuse Cloud no longer pays for image or video generation. The studio keeps the generated files on your device and uses the provider you selected. Four illustrated inspiration cards sit above the description on Android, iPhone, web and desktop. Tapping one fills its description and style; it does not replace the face until you ask your own provider to draw and then choose a result. The examples are Moss Courier, Tide Tinker, Dusk Moth and Copper Acorn. Their five-state image sets are under assets/avatar-presets/; generation details and the non-sensitive prompts are in source.json there. Names, descriptions and style hints follow the app language; the stable preset ids are not user-facing copy.
  3. Four to choose from. The same subject four times (classic colouring, lighter, darker with a small accessory, a playful take), in the studio's 3D 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.

The inspiration assets are labelled as AI-generated in their canonical source.json. They were made through authenticated, metered Bailian model-inference HTTP APIs, not Model Studio's experience UI: qwen-image-3.0-pro used the native synchronous multimodal endpoint for 20 pictures. The manifest records the endpoint type, maintainer attestation, prompts, per-file SHA-256 and official API and terms links, but no credential, account, private host, request id or signed result URL. The manifest also records the terms distinction and the GPL contribution of any rights held by the API caller; it is factual project documentation, not legal advice.

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 the [image] block in config.toml ([llm] image_model still works), the Making pictures row on Connections in the web console, or Settings → Models on the phones and the desktop; empty means Automatic: the chat model's provider when it draws (qwen-image-3.0 on Model Studio), else another image provider configured by the person, else none, 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 from a provider configured by the person, such as Wan on Model Studio, through the asynchronous video API (an upload, a task, polling, the MP4), the [video] slot (or the older [llm] video_model, configuration.md) 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.

Your own clip ​

A short video of your own can take the place of the clips. On the phone, pick one under the pen on the face in the profile (Your own clip…) or in the Your own clip section of the avatar studio; on the desktop, in the avatar studio or the profile drawer. The clip plays muted in a loop wherever the face moves: for the mood at rest, and for every other mood too, since there is only the one clip. Where a still is shown (under the clip until its first frame is on screen, in the share cards, in the tiles) the stills of a studio face keep their place, and with no studio face the clip's first frame is the still. Before anything changes, a preview plays the clip in the face's own round frame, with Use and Cancel; Default face removes the clip, and the studio face or the dragon shows again.

The limits, the same on every device (every-device.md): MP4 or MOV, at most 15 seconds and 30 MB, with H.264 video; HEVC too where the device can play it (iOS always; the desktop where the system has the decoder, and its preview refuses a clip it cannot play; Android takes H.264 only today). A longer or larger clip, or another format, is refused with a sentence that states the limits; nothing is trimmed or re-encoded, and the sound is never played.

The clip stays on the device that chose it. It is not uploaded to nanoMuse Cloud and does not sync to your other devices, unlike the stills of a studio face; each device chooses its own. On Android it lives under minis-global/nanomuse/avatar/own-clip.mp4 with its first frame beside it as own-clip.png, filed with the signed-in account's other files; on iOS under the app's Application Support/nanomuse/avatar/own-clip.mp4; on the desktop under $DSH_HOME/nanomuse/avatar/own-clip.mp4. It survives a new studio face and a reset to the built-in one.

On Android that folder is inside the app's own storage, which no file manager or shell can open, so the workbench's menu has Save to gallery: it copies the base picture, the mood stills, the four candidates of the last generation, the animated moods and your own clips to Pictures/nanoMuse/avatar and Movies/nanoMuse/avatar, each batch named by the minute it was saved, and says how many pictures and clips landed. Nothing goes anywhere but the gallery of the same phone; it is the way to keep a face or a clip before clearing the app's data or moving to a new phone. (#664)

GPL-3.0-or-later. nanoMuse is an independent community project, not affiliated with Meta.