API for AI avatar talking-head videos from a script: how to choose

It depends on what you are building. For a finished talking-head video made from a script, you want an avatar API: you send the text and an avatar id, the vendor voices the script and renders the presenter as an asynchronous job, and you collect a video file. HeyGen, Synthesia, Argil, Creatify and Sume Avatar 1.0 each document one. Choose between them on five points: whether you need a rendered file or a live avatar that talks back, which avatar you start from (stock, custom, or a photo), the languages your scripts are in, how long each video runs, and how the vendor bills (per minute, per second, or plan credits).

Vendor facts below come from each vendor's own developer docs and pricing pages, read on 2026-09-29 (links under Sources). Sume facts come from Generate avatar video, the Sume API reference and pricing code. Starting from a voice track instead of a script is a lip-sync job, covered below. For one-to-one comparisons, see Sume vs HeyGen and Sume vs Synthesia.

Which APIs turn a script into a talking-head video?

All five take a script in the create request and deliver the video asynchronously. The table is alphabetical.

Which one should I pick?

Pick by the constraint that matters most to your job. Each point comes from the vendor's own pages, not from a quality test.

• HeyGen, if you want pay-as-you-go API credits with no subscription, batches of up to 100 videos in one call, output above 720p, or a choice of engine per request: Avatar IV is the default, Avatar V is an explicit opt-in that needs an eligibility check, and Avatar III is its dedicated photo-to-video pipeline.
• Synthesia, if your team already builds from Synthesia templates or needs dubbing and translation through the same Video API. Test videos are free and watermarked, up to 30 a day.
• Argil, if you build a video as ordered moments, each with its own avatar and either a transcript or an audio file.
• Creatify, if you choose from 1,500+ personas, or also want its link-to-video ad endpoints.
• Sume Avatar 1.0, if the talking head is one shot in a larger job, next to video models, text to speech, lip sync, captions and Timeline in the same API.
• Sume lip sync, if you want to make the voice track first, as a Sume text-to-speech output, and animate a still photo or an avatar to it: see the routes below.

Do I need a rendered video or a live avatar?

If the avatar has to talk back to a user in real time, you need a live avatar product, not a script-to-video API: Synthesia's docs describe its Interactive Avatars API as "synchronous and streaming, not a video file", and HeyGen sells LiveAvatar on plans separate from its API plans. Every Sume route in this post is a job that returns a Sume-hosted file. Real-time AI avatar vs video avatar API compares the two kinds.

When is Sume Avatar 1.0 not the right pick?

When the script is not in English, runs past 60 seconds, needs more than 720p, or the avatar has to talk live. In current code, Avatar 1.0 speaks English only. For another language on Sume, run two jobs: POST /v1/tts-1.0/generate with the transcript and its language (the schema says to set it for every non-English transcript), then POST /v1/veed/fabric-1.0 to lip-sync a still or a ready avatar to that Sume-hosted audio.

Avatar 1.0 also needs a ready avatar. You create one with POST /v1/avatar-1.0/generate from a prompt, structured traits, or a public reference image, at $0.95 per avatar; Create a reusable AI avatar walks through it.

What does a production script-to-video pipeline need?

Rendering is asynchronous at every vendor, so the same plumbing applies. On Sume:

• Send an Idempotency-Key on each create and store the job id. A retry with the same key and the same body returns the original job instead of billing a second one.
• Take the result by webhook (mode: "webhook" with a public HTTPS webhook_url) and keep status polling as the fallback, because Sume sends terminal job events only.
• Split long scripts by scene: one Avatar 1.0 job covers 4–60 seconds, so a revised paragraph means re-rendering one job, not the whole video.
• Keep voice separate from the face when you need control over it: make the audio with text to speech, then lip-sync a still or an avatar to it.

Which Sume route fits which input?

Sume has three talking-head routes, and you pick by what you already have: a ready avatar and a script go to Avatar 1.0; Sume-hosted audio and a still go to lip sync (POST /v1/veed/fabric-1.0, or POST /v1/minimax/h3-max/lip-sync for 5–14.8 seconds of audio); a driving video goes to Kling 3.0 Motion Control (POST /v1/kling/3.0/motion-control). Each has its own guide: talking avatar video, lip sync and motion control.

A plain video model is not a fourth option: Sume's models guide says video models do not lip-sync to generated TTS or to a later voice-over.

When should I use the Avatar 1.0 talking video?

When you have words but no recording. Send the avatar's avatar_handle and exactly one of script or video_inputs; the talking avatar video guide covers quality, aspect_ratio and the optional inputs.

• To check first-frame stills before paying for the full render, create an avatar video preview, then call generate-video on its id.
• For captions, burn them onto the finished video with POST /v1/video-captions.

When does lip sync or motion control fit instead?

Lip sync fits when the voice track exists before the face: it animates a still or an avatar to that audio, with no script involved. The audio has to be on Sume's media host already, such as a Sume TTS output; the API refuses audio from other hosts. Each job takes exactly one visual source, image_url or avatar_id / avatar_handle, plus that audio_url. Fabric takes up to 300 seconds of audio per job, and H3 Max Lip Sync is the explicit alternative inside a 5–14.8 second window.

Motion control fits when the performance comes from footage. It animates a still with the motion of a driving video you supply: send motion_video_url (public HTTPS, at most 30 seconds), its length in duration_seconds, and one visual source. keep_original_sound keeps the driving video's audio track unless you set it to false, and prompt only steers appearance details. To put an avatar's face onto an existing video instead, Face swap is in Beta.

What if my video is longer or has two speakers?

Then the per-job limits decide the plan:

• Past 60 seconds of script, Avatar 1.0 needs several jobs: the docs say to split a longer script, and How to make an AI avatar video longer than 60 seconds joins the parts.
• Up to 300 seconds of finished audio fits one Fabric job, and up to 30 seconds of driving video fits one motion-control job.
• Two speakers need two jobs, because current Avatar 1.0 execution supports one resolved avatar per final video. AI avatar conversation video cuts the turns together.
• Silent pauses can stay inside one Avatar 1.0 job: a video_inputs scene with voice.type: "silence" is a non-speaking beat.

How do the prices compare?

Units differ, so convert before you compare. HeyGen quotes USD per minute charged by actual seconds, Synthesia and Argil bill plan credits, Creatify bills credits per 30 seconds or per second by model, and Sume bills per second. Sume's Avatar 1.0 rates work out to $11.04 standard, $14.70 plus, $33.00 max per minute (no product image), plus a 5.5% agent fee by default; the rate rises with a product image, and those rows are on API pricing.

On Sume, lip sync counts audio seconds and motion control counts driving-video seconds, each rounded up. To price a real job, multiply your script's spoken length by each vendor's unit, and add avatar creation ($0.95 per avatar on Sume) where you need a new avatar.