Digital humans: make a face talk, re-sync a video's lips

Digital humans aren't a separate area in Mosael — they are two combinations of inputs inside video generation:

	You give it	You get
Talking photo	A portrait + a voice-over	A video of that face speaking the voice-over
Lip-sync	An existing video + a new voice-over	The same footage with the person's mouth matching the new audio

The length follows the voice-over: these models take no duration setting, and the UI says "duration follows the driving audio". If the audio is longer or shorter than the model accepts, you're told when you submit ("the driving audio is 72 s; this model takes at most 20 s") instead of a job that is bound to fail.

Which models#

After you connect a provider, the models appear in every model picker — only the ones that declare talking photo or lip-sync.

Provider	Model	What it does	Audio length	Needs object storage
Alibaba Bailian	wan2.2-s2v	Talking photo (480P / 720P; checks for a clear frontal face first)	1–20 s	No
Alibaba Bailian	videoretalk	Lip-sync (each side of the source video must be 640–2048 px: anything outside is scaled into range first, so a 512×512 talking photo can go straight on to lip-sync)	2–120 s	No
Volcengine Jimeng	OmniHuman 1.5	Talking photo (720p / 1080p; an optional prompt steers motion and camera)	1–59 s	Yes
Kling	Kling Avatar	Talking photo (optional prompt)	2–300 s	No
Kling	Kling Lip-sync	Lip-sync (finds the face on screen longest first)	2–60 s	For the source video
HeyGen	Avatar IV	Talking photo (optional motion prompt)	Up to 10 min	No
HeyGen	Lip-sync	Lip-sync (mp4 / webm source)	—	No
Hedra	Character-3	Talking photo (aspect ratio picked from the image)	Up to 10 min	No

The providers that need object storage only accept public links, so local images and audio are uploaded to the object storage you set up in plugins first. All of these were built from the providers' official documentation; the first time you use one, keep an eye on the error message — it repeats what the provider said.

Where to use them#

• Character assets: "Make it talk" on the asset page — write a script, it is voiced with the character's own voice, and its front view (or the first image in pick order) becomes the talking photo. Character cells on boards have it too.
• Boards: "Make it talk" on image cells (pick a voice in the panel, or wire in an audio cell) and "Lip-sync" on video cells (audio from an upstream audio cell, or voiced from a script on the spot). The panel lets you pick the model; "Resolution" is under advanced options and lists only the tiers the chosen model has.
• AI Studio: pick a talking-photo or lip-sync model and attach the portrait / source video plus a driving audio.
• Agent: it knows how — voice first, then hand the face and that audio to a talking-photo model (generate_video).
• Workflows: the "Make it talk" (character / image), "Lip-sync", "Long script to segments" and "Dub lip-sync" nodes.

Templates#

The template library has three ready-made graphs:

• Script → talking presenter: one frontal face and a script. The script is voiced sentence by sentence, neighbouring sentences are grouped up to what the model accepts in one go (seams fall between sentences), each group is spoken by the face, the clips are laid end to end on the timeline, captions follow the measured voice-over, and the result is exported. A group shorter than the model's minimum (Kling starts at 2 s, and a lone "OK." can form its own group) is padded with a little silence; later groups shift accordingly.
• Video translation · lip-sync: after transcribing, translating and dubbing sentence by sentence, the speaker's mouth is matched to the translated voice-over. The source is cut into chunks the model accepts, at gaps between sentences; chunks without dubbing reuse the original for free. The fixed footage goes on a new top video track; the source is untouched.
• Product → on-camera presenter pitch: pick a character from the asset library as the host; it opens and closes on camera, with product shots in between. With a video model that takes a first frame, the beats in between move (as in Product into a narrated short, one more paid video per beat).

Consent#

A digital human needs the pictured person's consent. Mosael checks three things, the same way at every entry point (AI Studio, boards, the agent, workflows, scheduled tasks):

1. Every generation is confirmed: any generation with driving audio needs "I have the pictured person's consent" ticked — it's you, or they agreed separately. The agent has to pass that declaration too.
2. Character declarations: when the face is a reference image of a character asset that is a real person, it needs "This is me" or "The person has given consent" (fictional characters don't). Add it on the asset page.
3. Cloned-voice declarations: when the driving audio was voiced with a cloned voice, that voice must say whose it is (you, consented, or fictional). Voices created before this was asked are "undeclared"; declare them in the voice library before using them for a digital human.

Whatever is missing is named when you submit — which character, which voice. The agent says so when it opens the confirmation card, not after you approve it.

Cloned voices spoken through Bailian#

The driving audio can also be a library voice spoken by Bailian's CosyVoice: when you pick a voice for a character asset, on a board or in a workflow, choose the engine Alibaba Bailian (CosyVoice) and the list gains your library voices (the My cloned voices group in dropdowns, "Cloned · name" on asset pages and workflow nodes). What speaks is that voice's copy cloned into your Bailian account, with no local engine to install.

• Consent follows the same line: the voices that can be cloned on Bailian are the ones that can be used for a digital human — voices whose owner has been declared. Undeclared voices are never cloned and aren't listed.
• One more yes: the first time a voice is used with CosyVoice you agree, once, to upload its reference clip to your Bailian account. Voicing in AI Studio's Audio or in an audio cell on a board asks on the spot; a character's "Make it talk" and workflow nodes run in the background where nobody can be asked, so without that agreement they stop at the voicing step and say why — click Clone on Bailian on the voice under Settings → Dubbing & voices first.
• For cloning, copies that follow the key and the model, and deleting a voice along with its remote copies, see Editing → Your own voice through Bailian.

AI-generated label#

It is not just digital humans: an export that uses any AI-generated or synthesized media — generated images, video or music, AI voiceover, digital humans, and anything trimmed, denoised or separated from them (the "contains AI-generated content" flag follows an asset's sources) — gets labelled in the file. Muted audio clips don't count.

• In the file: the standard implicit-label (AIGC) metadata is always written; it doesn't touch the picture.
• In the picture: off by default. Tick Label AI-generated content "AI-generated" in the export dialog to burn it in: a large "AI-generated" in the top-left corner for the first 3 seconds and a small line in the top-right corner (white text on a translucent box) for the whole video, sized from the frame's short edge so it scales with the export resolution. The box starts unticked on every export; tick it for platforms that require an on-screen label. The workflow "Export timeline" node and a board's timeline cell have the same option, also off by default.

Cost#

Talking photos and lip-sync are metered by the actual length of the output (which is the voice-over's length). Bailian wan2.2-s2v (0.5 CNY/s at 480P) and videoretalk (0.08 CNY/s) can be pre-filled from their list prices in Admin → Cost rules; the other providers either show prices only after signing in or bill in credits, so enter your account's prices by hand. A lip-sync output follows the source video's size, so only a resolution the request actually states is recorded (it is no longer guessed as 720P), and the source video and driving audio are not counted as reference images; the duration in Bailian's response is checked against the recorded seconds.