AIM Digital Human Video Generator

Problem It Solves

When a user already has a portrait image and wants it to “speak” from a specified audio file or script, the workflow often involves TTS, base64 encoding, gateway submission, status polling, and secret-key placement. This skill collapses those steps into a repeatable pipeline: provide an image and either audio or text, then receive a directly usable video URL.

Core Workflow

• Input: local path to a person image; audio source can be one of: existing --audio-path, text plus voice sample --text / --voice-sample / --voice-sample-text, or text plus preset voice --voice-preset.
• Generation: if text is provided, TTS is first called through the AEP gateway; then the image and audio are read, converted to base64, and submitted to /video2, where the server returns a public TOS URL.
• Polling: ffprobe reads the audio duration, and the skill adjusts HTTP HEAD polling frequency by stage to reduce unnecessary requests.

Boundaries and Notes

• The secret key is written only to the skill root .env file, under the key aim-secret-key; a preflight check is required when it is missing.
• A single polling run defaults to a 60-minute timeout, but timeout does not mean failure; .task-history.jsonl is used for later re-checks.
• Action-description prompt is not exposed by default; when no audio and no voice preset are provided, the “Chinese female” preset is used.