Source
InfiniteTalk repository
Read the primary source.
Reference illustration, not a screenshot of this source.
InfiniteTalk
Open audio-driven avatar model · Video · Speakers
Turns a portrait or a clip and an approved audio track into a speaking video with head and body motion. Open weights from MeiGen-AI.
Use it for
The cheapest measured route for a portrait speaking our own audio: identity held, about $2.30 a published minute at 480p on WaveSpeed.
Choose something else for
Not for sealed lips on b, p and m in French from a small portrait: four variants failed Frank's eye on 14-09-2026.
Give it
Portrait image or clip; Approved audio; Optional prompt
Get back
Speaking video, 480p or 720p
How it is operated
Hosted API (WaveSpeed $0.03 a second at 480p) or open weights in WanGP on a rented card; measured 14-09-2026
Used in
Availability
Provider capabilities and historical Studio notes are distinguished below. Current account access, credits and runtime readiness were not tested.
How charging works
WaveSpeed $0.03 a second at 480p, $0.06 at 720p (read 14-09-2026); measured $0.12 for 3.1 s.
Where processing happens
Provider-hosted on WaveSpeed; self-hosted on a rented card.
Language support
French tested by us: identity and rhythm pass, bilabial closures fail at 480p.
Create a speaking presenter
★★★★☆ 4/5 for this task
The only born-whole route measured on our own speaker: identity held, about 70 s and $0.09 to $0.12 for a 3 to 4 s clip at 480p. Lip shapes on bilabials failed Frank's eye.
Decision: Start here for a moving speaker on approved audio; check the b, p and m frames before approving.
Limits and things to check
- The lips do not seal on the b of « de bord », even with the real voice, at 720p or with the word slowed (14-09-2026).
- The last word of a clip is under-articulated; a silent tail of about 1 s, cut afterwards, helps.
- Self-hosted, the 14B model ran out of memory on a 46.5 GB RunPod container on 20-08-2026.
Same family as LongCat-Video-Avatar: the whole person is generated from portrait and audio, not a mouth repainted on an existing clip.
Evidence
Measured on our own speaker 14-09-2026: six hosted clips, cost and wall clock read, judged by Frank
Catalogue review: 2026-09-15. The evidence note identifies what was verified. This date does not imply a fresh tool test or verification of every linked source.
