SadTalker is Apache 2.0: self-host it or call a hosted talking clip?
SadTalker turns one portrait and an audio file into a talking head. What it costs you to run, and the hosted Sume still-plus-audio route that skips the GPU.
SadTalker is the open-source project built around one idea, "single portrait image + audio = talking head video". Its README shows an Apache 2.0 license, along with a disclaimer against using it for fraud or deception. So the question for a small team is not whether you may run it, but whether you want to run it.
I read the SadTalker README on 2026-10-04. The Sume side comes from the models page and the code behind it.
What running SadTalker involves
The README lists a 256 px and a 512 px face model, a Still mode, and optional GFPGAN enhancement. It asks for Python 3.8 and CUDA 11.3, which you install and keep working yourself.
- You pin an older Python and CUDA stack and keep it working.
- You choose the face model size and decide whether to enhance faces.
- You store and serve the MP4s yourself.
The same inputs on a hosted route
Sume's POST /v1/veed/fabric-1.0 takes the same two things: a still (image_url, or a ready avatar's avatar_handle, never both) and an audio_url hosted on Sume media, up to 10 MB, with a duration_seconds that you measured. It returns a job. A completed job carries a media.sume.com MP4, which you can keep. resolution is 480p or 720p. A 5 second clip at 720p reserves $0.94 and a failed job is refunded.
H3 Max Lip Sync, POST /v1/minimax/h3-max/lip-sync, has the same body but only accepts 5 to 14.8 seconds of audio, with 480p, 768p and 1080p output, and a 5 second clip at 768p reserves $0.50. Pick it when the clip is short and you want 1080p.
A quick way to choose
Choose self-hosting when you already own the GPU, want the face model under your control, and can absorb the maintenance. Choose the hosted route when you have a few dozen clips a month and no GPU, because you pay per second of audio and nothing sits idle. If the audio does not exist yet, one request to POST /v1/avatar-1.0/talking-video takes a script instead, and runs 4 to 60 seconds.
Either way, check the photo's rights first. The SadTalker disclaimer and the same caution applies to any avatar tool: animate people who have agreed to it.
Sources
Related posts
More in Sume Avatar 1.0
- School principal weekly update: a 40-second avatar clip, 36 weeks
A 40-second weekly announcement from one AI avatar costs $7.36 at standard, or $264.96 for a 36-week school year. Includes a script frame and caption tip.
- Setup guide intro with a product image: +$0.01 to $0.03 a second
Adding product_image to a Sume avatar video raises the per-second rate by $0.01 on Standard, $0.013 on Plus and $0.03 on Max: +$0.30, $0.39 or $0.90 for 30 s.
- Split a 3-minute script into 60-second avatar jobs
Sume avatar talking videos accept scripts of an estimated 4-60 seconds. Split a 3-minute script into jobs of that size, then join the audio with timeline audio.
- Tavus custom faces need a paid plan: Sume creates one for $0.95
Tavus lists custom face creation on Starter, Growth and Enterprise plans only. See how that compares with creating a reusable Sume avatar for a flat $0.95.
Written by Sume