Multimodal
Lip sync
A dedicated POST /v1/lipsync endpoint
— not one more /v1/videos request
shape. A driving audio track re-syncs an existing face, in one of two modes depending on the model you
pick: image mode animates a still photo (image_url
+ audio_url), video mode
dubs an existing clip's mouth to new speech (video_url
+ audio_url). There's no
prompt field at all, and no
duration_secs either — output length
always follows the real length of your audio (and, in video mode, your source video too).
Lip sync works from your picture or video and your audio. If you'd rather pick a character from a
catalog and give it a script to speak (Arabic and 50+ other languages), use
Talking avatars —
POST /v1/avatars/videos.
This is deliberately kept off /v1/videos'
own start_image_url/input_video_url/input_audio_url
fields — those already mean something else there (starting frame, video editing, LTX's own audio-to-video),
and combining them for lip sync would be one field-name accident away from a reference-to-video call
(input_references/input_audio_references)
that generates something new instead of preserving your exact source face/scene and just re-timing the mouth.
Image mode: animate a still photo
import requests, time
job = requests.post(
"https://videorouter.sh/api/v1/lipsync",
headers={"Authorization": "Bearer llmr_sk_live_..."},
json={
"model": "fal/h3-max-lipsync",
"image_url": "https://example.com/face.jpg",
"audio_url": "https://example.com/speech.mp3",
},
).json()
# -> {"id": "fal:...", "status": "queued", ...}
status = requests.get(
f"https://videorouter.sh/api/v1/lipsync/{job['id']}",
headers={"Authorization": "Bearer llmr_sk_live_..."},
).json()
Video mode: dub an existing clip
import requests, time
job = requests.post(
"https://videorouter.sh/api/v1/lipsync",
headers={"Authorization": "Bearer llmr_sk_live_..."},
json={
"model": "fal/sync-lipsync-v3",
"video_url": "https://example.com/clip.mp4",
"audio_url": "https://example.com/new-speech.mp3",
},
).json()
# -> {"id": "fal:...", "status": "queued", ...}
status = requests.get(
f"https://videorouter.sh/api/v1/lipsync/{job['id']}",
headers={"Authorization": "Bearer llmr_sk_live_..."},
).json()
Real price is computed from your file's own probed duration — for video-mode models, that's
max(video_secs, audio_secs), never
the shorter of the two, since some models can stretch the shorter input to match the longer one. That cost
already includes our 2.5% platform fee.
Which models
12 models across 3 providers — fal.ai, Atlas Cloud and Vidu.
Image mode (image_url + audio_url)
fal/h3-max-lipsyncfal.ai (MiniMax)$0.05–$0.32/s by resolutionfal/sync-lipsync-v3-imagefal.ai (Sync Labs)$0.1333/satlascloud/infinitetalkAtlas Cloud (MeiGen-AI)$0.03/s (480p), $0.06/s (720p)fal/heygen-avatar4-image-to-videofal.ai (HeyGen Avatar 4)$0.10/s, billed in whole secondsfal/omnihuman-v2fal.ai (ByteDance OmniHuman 2.0, full-body talking video)$0.105/s of audio; audio 4–300 s; optional prompt for scene/cameraatlascloud/omnihuman-v2Atlas Cloud (ByteDance OmniHuman 2.0, same model, lower price)$0.0896/s of audio; audio 4–300 s; native 720p; optional prompt for scene/cameravidu/s1-avatar-offlineVidu (S1 talking avatar, lip sync + body motion)$0.005/s of audio, rounded up; 540p or 720p, same price; typically 1–7 minboson/higgs-avatarBoson AI Higgs Avatar (talking head from one face image; audio up to 60 s)$0.07/min at 640x480 / 480x640, $0.09/min at 640x640 (default), billed per second of audio, rounded up; usually secondsfal/kling-ai-avatar-v2-standardfal.ai (Kling AI Avatar v2 Standard)$0.0562/s of audio; optional promptfal/kling-ai-avatar-v2-profal.ai (Kling AI Avatar v2 Pro)$0.115/s of audio; optional promptwavespeed/kling-ai-avatar-v2-standardWaveSpeedAI (Kling AI Avatar v2 Standard, same model)$0.056/s of audio, rounded up to whole secondswavespeed/kling-ai-avatar-v2-proWaveSpeedAI (Kling AI Avatar v2 Pro, same model)$0.112/s of audio, rounded up to whole secondswavespeed/seedance-2.5-talking-avatarWaveSpeedAI (ByteDance Seedance 2.5 Talking Avatar)$0.22/s (480p), $0.44/s (720p, default); audio up to 120 swavespeed/skyreels-v3-talking-avatarWaveSpeedAI (Skywork SkyReels V3 Talking Avatar)$0.03/s (480p), $0.06/s (720p, default); audio up to 15 swavespeed/longcat-avatar-1.5WaveSpeedAI (Meituan LongCat Avatar 1.5)$0.04/s (480p, default), $0.08/s (720p)wavespeed/veed-fabric-1.0WaveSpeedAI (VEED Fabric 1.0)$0.07/s (480p, default), $0.14/s (720p)wavespeed/pruna-p-video-2-pro-avatarWaveSpeedAI (Pruna P-Video 2 Pro Avatar)$0.02/s (480p), $0.04/s (768p, default); audio up to 180 swavespeed/infinitetalk-fastWaveSpeedAI (MeiGen InfiniteTalk Fast)$0.015/s of audio, the cheapest talking-avatar rowVideo mode (video_url + audio_url)
fal/sync-lipsync-v3fal.ai (Sync Labs sync-3)$0.1333/sfal/sync-lipsync-react-1fal.ai (Sync Labs)$0.1667/s, max 15s in/outfal/sync-lipsync-v2-profal.ai (Sync Labs lipsync-2-pro)$0.0833/sfal/sync-lipsyncfal.ai (Sync Labs lipsync-1.9, the original)$0.70/min (~$0.0117/s)fal/sync-lipsync-v2fal.ai (Sync Labs lipsync-2)$0.05/sfal/veed-lipsync-v2fal.ai (VEED)$0.07/sfal/heygen-v3-lipsync-precisionfal.ai (HeyGen)$0.10/sfal/heygen-v3-lipsync-speedfal.ai (HeyGen)$0.05/sAtlas Cloud rows bill through Atlas Cloud's reserve-then-settle flow (same as every other Atlas Cloud video model) — the price shown is a pre-call estimate; the real charge settles from Atlas Cloud's own completed-job invoice. Every model here also shows up under the Lip Sync filter on the models catalog.
Boson Higgs Avatar. boson/higgs-avatar takes one face image and a driving audio file of up to 60 seconds; longer audio is refused up front with 400 audio_too_long and nothing is charged. Pick the output size with resolution (640x640, 640x480, 480x640) or aspect_ratio (1:1, 4:3, 3:4). Text-driven speech is not offered here, because the length of the result would not be known before the job runs. The same model also runs live inside your own LiveKit room: see Live video.
Not yet supported
Avatar/digital-human models that need a voice_id
or avatar selection (e.g. HeyGen's own Avatar-IV/Digital Twin) aren't onboarded — this endpoint has no
avatar/voice-selection concept, only a real image or video you supply yourself. Kling LipSync and Hedra
aren't wired in yet either. Self-hosted open-weight options (LatentSync, MuseTalk) would need their own GPU
deployment, not a simple resale row, and aren't on this platform today.