Game NPC dialogue with AI avatar clips: pre-render lines, not live

For an interactive story or a game with scripted dialogue, you can pre-render each NPC line as a short avatar clip. A bank of 120 lines at 5 seconds each is 600 seconds of video and costs $147.00 at plus, $110.40 at standard, or $330.00 at max. Sume avatars are async jobs: you submit a request, get a job id back, and read the finished video later. There is no live video session.

That works when the dialogue is written in advance. It does not work when a player can say anything and the character must answer in the moment.

Cost of a dialogue bank

Each line is a job of 4 to 60 seconds, so 5 seconds is a short but valid length. The table uses the no-product rates from provider-pricing.

Designing for clips

Branching dialogue fits clips well. Each node in your dialogue tree is a script, each script is one job, and your engine plays the clip when the player reaches the node. Keep lines short, because a 5-second clip plays and ends, and a 60-second one makes the player wait. Use a silence scene for a beat in which the NPC looks at the player without speaking.

Use the same avatar_handle for each character and a different handle for each character in your cast. Creating a handle costs $0.95, so a cast of eight is $7.60 once. Keep one scene prompt per location so the characters appear to share a world.

Production routine

Render in waves. Submit a wave of lines, poll the jobs, and review. Each paid generation can sit in queued until a worker takes it, and your plan sets the queue and concurrency limits, so a wave of 120 is better spread over several submits than fired at once. If the queue is full you get queue_full, and you wait for running jobs to finish and retry with the same idempotency key.

Render the first pass at standard and play-test. Re-render only the lines that need it, and only the final cast at max if you need that quality. A re-render of 20 lines at max is $55.00.

Where clips stop

If the NPC must listen to free speech or text and reply with new words, you need live generation of both the words and the video, and Sume avatars do not do that. A middle path is a fixed set of replies, chosen by your own logic from what the player picked or typed, and each reply is a pre-rendered clip. Players accept this in most narrative games, as long as the set is large enough that the replies do not repeat too soon.

A line manifest

Keep one manifest file with a row per line: a line id, the character handle, the script, the tier, the job id, the status and the media URL. The manifest is the source of truth. A small script reads rows with no job id, submits them with the line id as the Idempotency-Key, and fills in the job id. A second pass polls the open jobs and fills in the URL. If the script crashes, rerunning it is safe, because the same key returns the same job.

Using the line id as the key also gives you free deduplication: two designers who add the same line produce one render, not two.

Localization

Check the language support for your target before you commit to a localized cast. Avatar scripts that use captions need a caption style that suits the language: for Korean speech, the docs require a Hangul style and reject Latin-only styles with 400 caption_hangul_text_latin_style. Test one line per language first.