Avatars & voices

Avatars

An avatar is a likeness plus a baked idle loop (what it does while listening) and, optionally, baked gestures such as a wave. Each avatar is bound to an engine; the engine is derived from the avatar, so visitors never see a mismatch.

Built-in avatars

The repo ships a default portrait, a baked default avatar, and three FlashHead portrait presets (F, Nova and Luna) so the first run shows a face without any upload. FlashHead needs only a single portrait, which is why adding a preset is one file plus one line in internal/avatarcatalog. Nova and Luna are shipped as half-body presenters: FlashHead still renders a 512 face, and the worker pastes it back onto a 720×1280 head-to-waist canvas at the exact crop box (workers/avatar/halfbody.py builds the canvas, bbox and blend mask), so you get a half-body figure from a head-only model (Luna is the same recipe in a night-time indoor look). The half-body pipeline is prepare → recrop (loosen the crop so the whole head moves) → blend, then bake an idle from the console. The console lists them with their engine tag. Selecting one as current takes effect immediately for new sessions.

Bake your own from a video

1. Record 10–20 seconds of the person talking to camera: frontal, even light, mouth visible, no hands in front of the face. Phone video is fine.
2. 形象与声音 → 上传说话视频. The pipeline resamples to 25 fps, extracts frames, scores them for a closed-mouth frontal sharp "speaking base", runs face alignment with MediaPipe (no InsightFace, so the chain stays commercially clean), and bakes the engine-specific assets.
3. Watch the job in the same page; on success the avatar appears in the list, no restart.
4. Optionally bake an idle loop from the same material so the avatar breathes and blinks instead of freezing between answers.

Baking is a one-shot GPU job that needs 4–5 GB of VRAM. On a full pool it will fail with OOM; the console tells you to stop one worker first.

Consent. Training a likeness of a real person requires that person's permission. The upload form includes a consent confirmation and the policy is documented in Compliance.

Engines

Engine	Look	VRAM	Notes
wav2lipLS 384	half-body, crisp mouth	≈ 2 GB	lightest, runs on 6 GB cards
MuseTalk 1.5	half-body, smooth	≈ 7.7 GB	best tolerance to varied source material; bbox_shift knob for mouth opening
FlashHead	head-and-shoulders close-up, diffusion	≈ 6.6 GB	highest fidelity for close framing

Mixed pools are supported: one GPU can run MuseTalk for half-body avatars and FlashHead for close-ups, dispatching by avatar. See Hardware and the engine pool design.

Preload

Loading an avatar into a worker takes seconds. 预热 (preload) does it ahead of time for the avatars your channels use, so the first visitor does not stare at a black frame.

Gestures

Avatars can carry one-shot clips (today: wave). Visitors get a button; integrators get POST /action; agents get the mynah_action tool in the DeepSeek Harness plugin.

Voices

Which voices appear depends on the TTS tier in 配置中心 (see TTS tiers):

• EdgeTTS (default): hundreds of Microsoft neural voices. The list is cached and a featured set is pinned to the top (zh-CN-XiaoxiaoNeural, zh-CN-YunxiNeural, en-US-AriaNeural, …). Speed is adjustable; style instructions are ignored.
• Qwen3-TTS (local vLLM-omni): built-in voices plus voice cloning from a short reference clip. Upload in the Voices tab, the console transcribes the reference for you, and the new voice shows up in every dropdown. A per-voice seed keeps the timbre stable across sentences; style instructions ("更沉稳") are honored.
• Qwen realtime brain: when the conversation brain is set to Qwen realtime, speech is synthesized in the cloud and the voice dropdown shows the cloud voices instead.

Every dropdown has a 试听 button that synthesizes a sentence through the live configuration. Channels freeze the voice they were published with, so changing the console voice does not change a live channel until you republish.